What “AI agent” actually means when a vendor says it

Every piece of software sold to property managers is an AI agent now. The chatbot in the corner of a dashboard is an agent. The pricing tool that emails you a suggestion is an agent. The thing that writes your guest replies is an agent.

They are not the same product, they don’t cost the same, and they don’t do the same amount of your work. But they are all sold with the same word, which makes it very hard to compare what you are actually buying.

There are three genuinely different things underneath that label. Once you can see the difference, most vendor conversations get a lot shorter.

One: it answers

The simplest version is a system you ask questions and get answers from. A chat box on top of your data.

“What was my occupancy in August?” “Which properties are behind last year?”

This is useful. It replaces the fifteen minutes you would have spent building a report, and it means the answer is available to anyone on the team rather than only the person who knows where the export button is. It is also now the most widely available layer: PMS platforms are building their own assistants, MCP connectors expose your data to general-purpose models, and support bots can be pointed at the same tables.

What it doesn’t do is change anything. You still have to decide what to do with the answer, and then go and do it. And because you will act on what it tells you, you have to trust it completely — if the underlying data isn’t clean and structured, an answering system will confidently hand you wrong numbers, and you will make decisions on them.

The honest test: after using it, is your to-do list shorter, or do you just have better information to work through the same list?

Two: it recommends

The second version watches something and tells you what it thinks you should do. Your rates are below market for those dates. This listing has an occupancy gap. That owner’s revenue is down year on year.

This is where most of the category is trying to get to, and it is where the language gets slipperiest — because “identifies”, “suggests”, “flags” and “recommends” all sound like work being done, and none of them are.

A recommendation is not a decision, and it is definitely not an action. Somebody still has to read it, judge it, and carry it out. If a system produces forty recommendations a week and you action twelve of them, the other twenty-eight are not time saved. They’re a queue.

This is the part nobody says out loud: a tool that recommends adds work before it removes any. Sometimes that trade is worth it. Often the recommendations quietly stop being read after the third week, which is why so many operators have a pricing tool running on defaults.

The honest test: does it hand you a number, or does it change the number?

Three: it acts

The third version does the job. It looks at the same information, decides, makes the change in the system where the work actually lives, and tells you afterwards what it did.

This is what “agent” means in the sense the word was borrowed from — someone acting on your behalf, within the authority you gave them. A letting agent doesn’t send you a list of suggested tenants and wait. They do the work and report back, inside the brief you set.

Three things have to be true for this to be real, and all three are worth asking about specifically:

• It writes back. Not into its own dashboard. Into your PMS, your channels, the places your portfolio actually runs on. If the output is a screen you have to copy from, it recommends.
• It works inside your instructions. Floors, protected dates, the weeks you don’t want touched, the properties you’ve paused. Not as a settings matrix you maintain, but as rules it holds and respects.
• It reports. What it did, what it left alone, and why. A system that acts without an audit trail isn’t an agent, it’s a liability — especially when the change it made is one you have to explain to an owner.

The question that settles it

You don’t need to understand the technology to tell these apart. You need one question:

Ask a vendor that question and ask to see the answer for a real portfolio. Not a demo environment. A week of what it actually did, including the days it decided nothing needed doing.

None of this means one is better

An answering system is the right purchase if your problem is that nobody on the team can get to the data. A recommending system is right if you have the hours to action recommendations and you want a second opinion on your judgement. An acting system is right if your problem is that the work isn’t getting done at all — or can’t be done properly by hand.

Revenue management is the clearest case. If you manage more than a handful of properties, it is not humanly possible to watch every competitor and the market every day, check every open night on every calendar, and adjust rates with the research behind each change. So most of it runs on gut feeling, or simply doesn’t happen.

Most property managers we talk to have the third problem and have bought one of the first two. That’s not because they chose badly. It’s because all three were described with the same word — and because AI products are routinely oversold.

Where Revzy sits

Revzy is an AI revenue manager for short-term rental portfolios, and it sits in the third category: it takes over the tasks that are humanly impossible, time-consuming or tedious but still matter — pricing every night, tracking comp sets, clearing last-minute gaps — inside the instructions you give it, and tells you each morning what it did. You brief it in plain language, the way you would a colleague. If you want to see what that looks like on a real portfolio, watch a Tuesday.