Agents that act.

Lede

There is a test that separates the two kinds of artificial intelligence a hotel is currently being sold, and it takes one question. If you switched it off this evening, would anything stop happening?

The difference that matters

A chat assistant waits. It sits behind a button, it is very good at answering, and its output is text on a screen that a person then has to do something with. Switch it off and nothing stops; the work simply goes back to being done the way it was done before, slightly slower.

An agent is attached to a surface and triggered by an event. Its output is not text. Its output is a change in the state of the system: a message triaged, a task created, a draft written and left waiting, a flag raised on a figure. Switch it off and something visibly stops.

That distinction sounds academic until you watch it play out on a working day. A chat assistant relocates effort: the manager now has to know what to ask, when to ask it, and what to do with the answer. She has become the retrieval engine for a system that was supposed to help her. An agent removes a step that was going to happen anyway, and it removes it before she arrives.

Treema ships three agents. They are deliberately not one agent with three personalities, because they sit on three different surfaces and are triggered by three different kinds of event.

GEX, at the inbox

GEX sits on the Smart Inbox and is triggered by an inbound guest message.

What it does, in order: reads the message, separates what the guest is asking for from how the guest feels about asking, pulls up that guest's profile and history so the reply is not written to a stranger, retrieves the relevant policy from the property's knowledge base, drafts a reply with that policy cited, and marks the ones a person needs to answer personally.

What it does not do is send. That is not a limitation waiting to be lifted; it is the design. A drafted reply that is wrong costs an editor a moment. A sent reply that is wrong is a commitment the hotel has made in writing.

The output of GEX is therefore not a conversation. It is a queue that has already been sorted and partially answered, so that the first person to open it is editing rather than triaging. On a busy morning that is the difference between starting work and starting to find out what the work is.

LORA, at the shift

LORA sits on Shifts, Tasks and Internal Chats, and it is triggered by the ordinary events of an operating day: a shift ending, a thread in which somebody names a problem, a checklist that was supposed to be finished and is not.

Its output is work with an owner. A message in a housekeeping channel saying the vacuum on four has died is a message; after LORA it is also a task on the engineering board, with the floor on it, the thread linked, and the maintenance procedure attached. Nobody typed it in twice. Nobody has to remember it at half past ten tonight.

At the end of a shift it produces the briefing — not a summary of chat, which would be unreadable, but a summary of what changed: opened, closed, overdue, escalated, and which guest threads still end with a guest. The outgoing manager edits that rather than composing it, and adds the one thing no system could have known, which is her judgement about the people and the guests.

This is the agent that is least impressive in a demonstration and does the most in a month. Nothing it produces is clever. All of it is work that a competent supervisor would have done if they had been sitting still, which in a hotel they never are.

SAM, at the numbers

SAM sits on the workspace and on Orders, and it is triggered by movement.

Its job is narrow on purpose: watch the lines the property already tracks and say something when one of them moves in a way the preceding weeks would not have predicted. Not a verdict, not a recommendation, and above all not a target. A flag, with the movement described plainly and the figure it moved on named.

Two disciplines matter here. The first is that some of these figures are estimates, they are labelled as estimates in the product, and SAM does not launder that label away by mentioning them without it. The second is that a flag is an invitation to look, not a conclusion. The hotel knows things the system does not — that a wedding party checked out, that a road was closed, that a supplier changed a price in AED and nobody updated the storefront. SAM's value is that it notices the movement on the day it happens rather than at the end of the month.

What may act alone

Once agents can act, the interesting question stops being what they can do and becomes what they may do without asking. It is worth writing that boundary down, because a system that has not written it down has drawn it by accident.

• Read, classify, route, retrieve. Alone. There is no failure mode here worse than a message in the wrong queue, and a person will move it.
• Open a task, attach a procedure, put it on a board. Alone. A task created in error is deleted in a second; a task never created is discovered by a guest.
• Assemble a briefing. Alone, and then edited by the person who is accountable for the shift.
• Reply to a guest. Never alone. Draft, cite, wait.
• Anything involving money. Never alone, and never quietly.

The pattern is simple enough to state as a rule: an agent may do anything that is cheap to undo, and nothing that is a promise.

The flow underneath

None of the above works without a flow to hang it on, which is the part that gets left out of most conversations about agents. An agent with no defined trigger, no branch and no named recipient is a very expensive way of producing opinions.

The flow is the boring skeleton: this event, that condition, this task, that person, and this is what happens if nobody touches it in time. Treema ships 16 flow templates because most properties want the skeleton before they want the intelligence, and because the templates make the argument concrete — you can read one and disagree with it, which is more than you can do with a promise about automation.

Ask a vendor what their agent does when nobody is looking at it. If the answer is nothing, it is a search box with manners.

The workday test is the honest one. Lena arrives at seven. Under the old arrangement her first hour is discovery: the pad, the group chat, the person leaving, the board. Under the new one she opens a briefing that is already written, an inbox that is already sorted with drafts waiting, and a board where last night's problems are already tasks with names on them. She still makes every decision that matters. She just no longer spends the first hour of her shift working out what they are.