How do you measure ROI on AI for real estate?

Badly, in most cases, because the measurement gets designed after the deployment rather than before it. By then the baseline is gone and every number is arguable.

The single most valuable thing you can do is capture your current state before anything changes. It takes an afternoon and it is the difference between knowing whether this worked and having an opinion about it.

The baseline worth capturing first

• Total inbound inquiries for a month, from every channel including the ones nobody owns.
• How many received a genuine reply, and how many received nothing.
• Time from arrival to first reply, split by hour of day and day of week.
• How many reached a booked tour, and how many of those tours happened.
• How many follow-up attempts a typical non-converting inquiry received.

That last one is usually the most revealing and the least recorded. Most operators discover their average is around two, and that their converting inquiries needed more.

What to measure afterwards

The same things, which is the point of writing them down first. The comparison is only meaningful if both sides are counted the same way.

Be careful with response time as a headline. It will improve dramatically and it is the least interesting number, because speed is only worth something if it converts. Watch the chain instead: inquiries answered, tours booked, tours attended, leases signed.

The attribution problem, stated honestly

Leasing volume moves for many reasons. Seasonality, pricing, what else is on the market, how many units you had available, whether you ran an ad. Isolating the effect of one change is genuinely hard and anyone claiming a precise figure is overstating their confidence.

The more defensible measures are the ones closest to the mechanism. Inquiries that received no reply is a clean number that should go to near zero, and it is not confounded by the market. Tours booked outside office hours is another, because that volume did not previously exist.

The cost side

Count the whole thing: the invoice, your build time, the data tidying, and the weekly review. Then set it against recovered inquiries at your own conversion rate and your own value per lease.

If the honest arithmetic does not work at your volume, that is a real answer and worth having early. A small operation with low inquiry volume and good coverage may simply not have the gap this addresses.

Give it long enough to mean something

One practical warning about timing. Leasing is seasonal in most markets, and a month of data compared against a different month of data is mostly measuring the season.

Where possible, compare the same period against the previous year as well as against the month before. Where that is not possible, be explicit that seasonality is in the number rather than quietly attributing all of it to the change you made. That honesty costs nothing and protects you when somebody senior asks a hard question about the figures.

The numbers worth reviewing every week

Separately from the return question, there is a short list worth looking at while a deployment is still settling.

• How often the agent handed off to a person, and whether the reasons cluster.
• Any conversation where it said it did not know, since those are gaps in what it was given.
• Bookings that produced a no-show, which can indicate qualification is too loose.
• Anything a resident or prospect complained about, read in full rather than summarised.

Those are not return measures, they are quality measures, and quality is what the return depends on. A deployment that is booking plenty of tours while annoying a fraction of callers is not actually succeeding, and the headline numbers will not tell you that.