Four different products get sold to you as AI

Three vendors call you in the same month. All three say the word AI. The first sells something that writes your owner emails. The second sells a box on your website. The third sells something that answers your phone at two in the morning.

Those are not three versions of one product. They are three different purchases, and the third one is governed by federal law in a way the first two are not.

So here are the four things the word is covering, sorted by how much can go wrong. The useful question turns out not to be how clever any of them is. It is who the thing is allowed to talk to, and what it is allowed to change.

The first one writes a draft and a person still presses send

This is the most common by a distance. When Buildium asked 1,796 property management professionals what they used AI for in its 2025 industry report, the top two answers were generating messaging to use as a starting point for customer communications, and writing property descriptions for rental listings.

Both produce something a person reads before anybody outside your company sees it. That single fact is why they went first. If the draft is wrong, you notice and fix it, and the cost of the error is thirty seconds.

It also sets the ceiling. A drafter saves you the blank page. It doesn't save you the reading, and if your team is not actually reading, you have bought a faster way to send the wrong thing.

The second one talks to a stranger in your name

A website chat box is the third use on Buildium's list, and it is a real step up in exposure, because now something is speaking to a person who does not work for you.

Two questions separate a good one from a bad one, and neither is about the technology. Is it working from your actual information, meaning your units, your rents, your pet policy, or is it improvising from a general model? And how fast does it hand a person the conversation when it doesn't know?

The second question is the one that gets skipped in the demo, because in a demo it always knows.

The third one answers the phone, which is measurably the hardest

Voice is not text with a speaker attached. It is a harder problem, and the size of the gap is the least understood number in this whole category.

We went through the benchmarks in detail in an earlier piece. The short version: on published customer service tasks, strong text models complete around 85% of them. Voice agents manage between 31% and 51% under clean conditions, and less once you add background noise and a range of accents. Speech recognition takes a cut. So does an accent, a dog barking, and holding onto who somebody is across four turns. All of it happens before the model ever reaches the question.

None of which means don't buy it. It means the demo you sat through was a clean line in a quiet room. Your worst call is a woman on a highway with a crying child.

If it calls your tenant rather than answering, federal law applies

This is the distinction almost nobody selling you a voice product will draw, and it is a bright line.

On 8 February 2024 the Federal Communications Commission released a Declaratory Ruling confirming that AI technologies generating human voices count as an artificial voice under the Telephone Consumer Protection Act. The consequence is stated plainly in the ruling itself.

As a result, calls that use such technologies fall under the TCPA and the Commission's implementing rules, and therefore require the prior express consent of the called party to initiate such calls absent an emergency purpose or exemption.

FCC Declaratory Ruling 24-17, CG Docket No. 23-362, adopted 2 February 2024

So the direction of the call changes the product. A tenant ringing your office and reaching a machine is one thing. That same machine ringing two hundred tenants about a rent reminder is a different thing, and it needs consent you can show you obtained.

The ruling was aimed at voice cloning scams rather than at property managers, and we have found no enforcement action against a maintenance reminder. That isn't the same as it not applying. Ask any voice vendor how their outbound calling handles consent. Watch whether the answer is specific.

The fourth one changes something, and that is a different category

The last of the four is the one the industry currently calls an agent. It doesn't produce words for you to check. It opens the work order, assigns the vendor, sends the appointment, updates the record.

The risk here is not that it says something wrong. It is that a small error becomes four more actions before anybody looks. A benchmark called tau-bench measures precisely this by running the same customer service task eight times and asking how often the agent gets it right on all eight. Leading models scored under 25% on retail tasks.

Read that as a statement about consistency rather than competence. It usually works. Usually is a fine standard for a draft and a poor one for something touching your ledger unsupervised.

This category is also the one the states have started writing rules for. California's privacy regulator has adopted rules on automated decisionmaking technology in which a significant decision expressly includes housing, defined as any building used or intended to be used as a home. A business using such a system for a significant decision must be in compliance no later than 1 January 2027. The obligations are a plain-language notice before you collect the information, and an opt-out, unless you offer an appeal to a human reviewer with authority to overturn the decision.

Note what that last clause does. It writes the human reviewer into the regulation as the price of not offering an opt-out. There is a carve-out worth knowing too. A system that provides or denies housing purely on whether a unit is vacant, or on whether payment came through, isn't making a significant decision.

Gartner counted roughly 130 real ones out of thousands claiming it

Enough vendors have noticed which of these four sounds most impressive that there is now a name for the practice. Gartner calls it agent washing, meaning existing chatbots and automation rebranded as agents, and in a June 2025 release estimated that of the thousands of vendors claiming agentic AI, only around 130 genuinely have it.

You don't need to adjudicate that. You need one question, and it works on all four at once. What does this do without a person in the loop, and where exactly does it stop and wait?

A vendor who can answer that in one sentence has built the thing. A vendor who answers it with the word intelligent has not.

Where we stop, so you can compare it to anyone else

Our own product drafts, dispatches, chases and logs. People do the physical work. Anything legally sensitive waits for your approval before it leaves the building.

We put that in writing because it is the answer we just told you to demand, and it would be a strange article that asked the question and dodged it.

Sort the pitch before you judge the price

Next time somebody says AI on a call, the useful move is not to ask what model it runs. Ask which of the four it is, then ask who it talks to and what it changes.

The answer tells you what a mistake costs you, and that is the number the price has to be weighed against.