Do the drafting first. Do the screening last.

Adoption in this industry went from 20% to 58% in one year. The share of companies that fully automated any process stayed at 8%.

Both figures are Buildium's, from the same survey. Almost everybody started. Almost nobody finished.

The usual explanation is that the technology disappointed. A likelier one is that people began wherever the last sales call pointed them, and vendors point hardest at the places where being wrong is expensive. So here is an order, with the reason for each position. The reason is always the same thing. What does it cost you when this gets one wrong?

Before any of it, write down how you actually decide

Two operators who have done this in front of audiences say the same thing, and neither one sells AI.

Don't worry about the AI and the zaps and all the those sexy bells and whistles. It's policies and procedures and your processes that, you know, if you don't have rock solid policies that are then put in action with your processes, all AI is going to do is screw it up even more and frustrate people.

Wolfgang Croskey, on the Triple Win podcast

Jo Oliveri, who builds workflow software, puts the dependency in four words on DoorGrow's podcast: process is built on policy. Her reason is worth the sentence. Policy creates the protocol where you can make decisions, and an automated workflow is only the logic you created through your process, running faster.

There is survey evidence for the same point from the other end of the industry. NAREIM asked 72 professionals across 38 real estate investment management firms to rate themselves in its 2026 technology survey, run with Juniper Square. They put their own AI maturity at 5.7 out of 10, their data quality at 6.2, and their governance readiness for AI at 5.1. The summary is one sentence: AI adoption has outpaced data quality, governance, and talent readiness.

Those are institutional firms with technology budgets, rating themselves below halfway on being ready to govern the thing they've already deployed. It's a small sample and it isn't your business. It's also the direction everything else points.

The practical version is unglamorous. If your late fee policy lives in three people's heads and those three disagree, no vendor is going to fix that for you, and the fastest way to find out is to try writing it down this week.

First: the things where being wrong costs thirty seconds

The three uses property managers reported most in Buildium's earlier survey of 1,796 professionals were generating a starting point for customer messages, writing listing descriptions, and answering website questions with a chat box.

That is not timidity. Two of those three produce a draft that a person reads before it leaves the building, so a bad output costs you the time it takes to notice. Start there. It teaches your team what the thing is actually like, on work where failure is embarrassing rather than expensive.

It also gives you an honest read on your own people. If nobody's reading the drafts after two weeks, you have learned something important about the next four steps, and it cost you nothing to learn.

Second: intake, because the volume is real and the judgment is thin

Maintenance intake is the one place in this business where the work is genuinely high volume and genuinely low ambiguity. A tenant describes a problem. Somebody categorizes it, decides how urgent it is, and starts it moving.

At 3.4 requests per unit per year, a 300 door portfolio produces about four of these a working day, every day, forever. That is the shape of task worth handing over, and it is why the vendors with real deployments in this industry landed on it rather than on anything more impressive.

One condition, learned the expensive way by somebody else. Brian Teeter, a property manager in Little Rock, had a tenant whose request had been denied simply call it back in through the AI and get it dispatched. His response was to make human approval a required step before any dispatch. Put that gate in on day one, not after your first incident.

Third: chasing, which is the largest thing nobody counts

After intake comes the least glamorous category and the biggest one. The vendor who has not confirmed. The tenant who has not replied about Thursday. The owner approval sitting unanswered since Friday.

This sits third rather than first because it only works once the two before it are running. Chasing requires knowing what was promised and by whom, which is the record that intake produces and the policy that step zero defined.

Nobody has ever measured how many hours a week this takes, in this industry or anywhere near it. We looked. It is the largest unmeasured cost in the business.

Which is a reason to instrument it yourself before you buy anything for it. Count the follow-ups your team sent last week. That number is your baseline and nobody else has one to sell you.

Fourth, and carefully: anything that calls or texts a tenant

Outbound is where this stops being an operations decision and starts being a legal one.

In February 2024 the Federal Communications Commission ruled that AI technologies generating human voices are an artificial voice under the Telephone Consumer Protection Act. Calls using them therefore require the prior express consent of the called party, absent an emergency purpose or exemption, on residential lines and cell phones alike.

Answering a call your tenant placed is a different act from placing one to your tenant, and only one of them turns on consent you have to be able to show. The ruling was aimed at voice cloning scams and we've found no action against a property manager. That isn't a reason to skip the question.

Last: screening, and there is now a date on it

Screening goes last because it is the only item on this list where a mistake can end in federal court, and because the record already exists.

In Louis v. SafeRent Solutions in the District of Massachusetts, tenants alleged that a screening algorithm produced disproportionately low scores for Black and Hispanic applicants and voucher holders, in part by not counting the voucher subsidy. The United States filed a statement of interest in January 2023 saying the Fair Housing Act applies to algorithmic screening. The case settled for $2,275,000 with final approval in November 2024, and SafeRent agreed for at least five years not to issue approve or decline recommendations from its score unless the model has been validated for fairness by civil rights experts.

California's privacy regulator has since written rules that name housing directly. Under the regulations, a significant decision includes the provision or denial of housing, and a business using automated decisionmaking for one must comply by 1 January 2027. It owes applicants a plain-language notice before it collects their information, and an opt-out, unless it offers an appeal to a human reviewer with authority to overturn the decision.

If you operate in California that is a deadline. If you don't, it's still the clearest description anybody has written of what a defensible screening process looks like, and a regulator wrote it rather than a vendor.

Expect the first attempt to go badly, and budget for that

One more thing belongs in the plan, and it is the part nobody puts in a proposal.

NARPM's own podcast ran an episode in March 2026 called Mistakes I Made Implementing AI. The guest is Inaas Arabi, chief operating officer at Block and Associates Realty, and the show notes describe a failed implementation attempt and what he learned from it. A trade association doesn't schedule that episode for a technology that installs cleanly.

The reason to sequence this at all is that each step teaches you something the next one needs. Drafting tells you whether your team will actually supervise. Intake tells you whether your categories hold. Chasing tells you whether your record is trustworthy. By the time anything touches an applicant, you have four rounds of evidence about your own company, which is more than any vendor can give you.

The order is the strategy

There is no version of this where you buy the impressive thing first and work backwards to the policy it needed. The 8% is what that looks like at scale.

Start where a mistake costs thirty seconds. Earn your way down the list.