How to make a property management company AI-native: a playbook

Onsite apartment staff turn over at 29.2% a year, against 14.2% for corporate staff[42]. Administration and payroll cost $2,323 per apartment in 2024, nearly 20% more than in 2021[41]. Each departure takes knowledge with it: which fees apply, which vendor fixes the old boilers, what the owner agreed to spend.

So operators are buying AI to answer the inquiries, take the maintenance requests and chase the rent. Nearly 80% of NAA survey respondents say it has helped, but most call it "moderately helpful", and in 2024 more than half were still waiting for industry guidance on how to govern it[43].

Meanwhile the sector's defining AI case so far is an antitrust case about rent. The Department of Justice and state attorneys general sued RealPage and large landlords over pricing software in 2024, and the proposed settlement reads like a governance rule: a landlord sets the limits, and a person can always override the machine[22]. Separately, the FTC has settled with Invitation Homes for $48 million over fees and deposits, and with Greystar for $24 million over fees[6,8].

Those cases sit exactly where agents are now being switched on: quoting rent, screening, pricing renewals, charging deposits. The hard part is answering every message right, under the rules of each property, and being able to show who decided what.

A useful test: could the firm show a resident, an owner or a regulator, for any message last month, which approved facts it used and which person was responsible for the decision behind it?

A property manager will add agents around its property management system (PMS), not replace it. The PMS may bring its own leasing agent; the maintenance bot, the screening service and the pricing tool come from other vendors. Each can govern its own agent. None answers the questions that cross them: is this quote from the approved fee schedule, which person is this renewal offer going out under, and where is the one record of it all.

OrchKernel is built to be that layer. Agents ask it before they act. It checks the action against the firm's rules, holds it for a named person when needed, runs it on that person's authority, and writes it to a tamper-evident log. What it is not: it is not a PMS, a screening service, a pricing tool or a fair-housing classifier. The PMS stays the system of record for leases, ledgers and trust accounting, and OrchKernel governs only the actions agents send through it. The detail is in the OrchKernel blueprint.

A rental unit goes around the same loop every time a resident leaves. Software can already do part of every step. At four of them, a mistake costs a renter a home, a deposit or a fair price, and those are where the law has been enforced.

The best public window into property-level costs is the NAA, IREM and BOMA Income/Expense IQ, which covers more than a million units. Its 2024 same-store figures per apartment[41]:

NAA calls this "the cost-efficiency phase of the operating cycle." Labor is the pressure point: 49% of operators with 5,000+ units name staffing as their single top challenge[42], and more than 70% of operators with 10,000 to 29,999 units rate labor conditions as significantly disruptive, against 39% of those under 1,000 units[44].

No independent study measures time by task in property management. The likely places are after-hours inquiries, maintenance intake and vendor chasing, move-out statements against a state clock (with photos, now, in California[27]) and month-end owner reports. Measure your own before you switch anything on.

Third-party managers earn a fee on revenue. One disclosed multifamily fee is 4.0% of gross revenues for management, leasing and construction supervision[5]. That is a single data point, not a benchmark. At 4% of $21,017, a manager earns about $840 per unit per year, while the property spends $2,323 per unit on administration and payroll. In many management agreements, onsite staff are a property expense the owner reimburses (to be confirmed against your own agreements).

So AI that removes onsite hours mostly lowers the owner's operating expense. Owner-operators such as REITs keep the whole gain, which may be part of why they move first. A third-party manager keeps it only through its own overhead, more units per manager, or new pricing. Expect owners who see payroll fall to ask for lower fees or a share of the savings; we found no public data on how this is being negotiated, so raise it before they do.

BLS counts 460,400 property, real estate and community association managers; 31% are self-employed[1]. AppFolio alone serves 22,096 property management customers with 9.4 million units, about 425 units each on average (our arithmetic)[4]. At the other end, Greystar manages 1,014,091 units[47]. Ownership splits the same way: individual investors own 58.8% of rental homes in buildings of two to four units, but 4.6% in buildings of 50 or more[48]. The median firm is small, runs one PMS and has no data team, which is why this playbook starts with fact sheets rather than models.

Vendors are named because the sources name them, not as recommendations. Figures from vendors are their own claims. The last column is what matters for governance.

In NAA's member research, leasing and resident communications showed the most impact in 2025; maintenance and reporting are growing; rent collection appeared in 2026[43]. More than half of 2025 respondents used general-purpose public AI tools, which raises the question of what resident data went into them. NAA does not publish sample sizes in its summary, and we found no independent adoption rate.

Owner-operators now describe AI in their annual reports. AvalonBay reports using AI "in correspondence with prospective, current and prior residents", and warns that its outputs "may be incomplete, inaccurate, misleading or otherwise flawed, and such issues may be difficult to identify or detect"[2]. Invitation Homes reports "an AI leasing assistant"[3].

Zillow published a fair-housing guardrail in 2024, a stop list and classifier aimed at steering[59]: the kind of check a leasing agent's replies need.

Into software for operators. EliseAI raised $350 million at a $4 billion valuation in September 2026, with annual recurring revenue past $200 million and software TechCrunch reports is used in one in six US apartments[55]. AppFolio grew revenue 20% to $950.8 million in 2025[4]. Thoma Bravo bought RealPage in 2021[60].

Tech-forward management companies exist. Belong manages single-family homes with the line "Powered by AI. Made and Loved by People."[57], and Doorstead raised $21.5 million in 2023[56]. But we found no AI-native manager with meaningful multifamily scale. The capital went to the software that existing managers license, priced per unit.

So a management company gets there by changing its own operating model with tools it buys.

Six stages, ordered by what an agent touches: first facts, then messages, then owner money, then renters' housing and rights. The order follows where operators already report results (front office first, then maintenance, then rent collection) and where enforcement sits. Durations assume a firm of 1,000 to 20,000 units; different properties can sit at different stages, and a HUD-assisted property may stay at Stage 3 for notices long after a market-rate one moves on.

Stage 0 for a handful of properties, then two Stage 1 workflows and a careful start on Stage 2. Pick inquiry replies and maintenance intake: both are high volume, neither decides anything about a renter's housing, and both produce numbers you can compare within the quarter.

At day 90, two questions matter more than the number of conversations handled. Did any quote go out without its fees? And can you show an owner, for each of their properties, every message the agents sent and every approval behind them?

We found no survey of how roles are changing. Where a filing or trade source exists, it is cited; the rest is our inference from the stages.

An agent can prepare each of these: gather the file, draft the letter, propose the number. A named person makes the decision, and the record shows whose it was.

The firm is the party in front of the renter. A property manager acts as the owner's agent and is usually the one that advertises, screens, prices, collects, holds deposits and serves notices. None of the rules below exempts a step because software took it.

Federal agencies are pulling back in places while states move forward. HUD has proposed dropping its disparate-impact rules and tried to revoke its 30-day notice rule, while California brings housing decisions under its automated-decision rules by 2027 and Colorado has passed an AI law that names housing[14,33,34,35]. A firm with properties in several states should plan to the strictest rule that applies to each property.

Most of this enforcement predates generative AI, and it hits what agents are now asked to do.

These are scenarios, not reported cases. Each is something an agent could do at a property manager, and each maps to one of the control points.

Scenario E, done right: the offer reaches the resident only inside limits a named person set, or with that person's decision on record.

This is where the controls in the rest of the page land. Each control point below has an owner at the firm (a property manager, the revenue manager, the broker of record) and a mechanism that enforces it when an agent acts through OrchKernel. Actions an agent takes directly in the PMS or a vendor portal, outside OrchKernel, are not governed by it, and the PMS remains the system of record.

Where a property manager needs a control whatever tools it runs, who owns it, and what enforces it.

OrchKernel is source-available under the Business Source License 1.1 and runs on your own servers, so you can read the code that enforces these controls. The terms are on the license page.

Record the baseline before Stage 1 and track the same numbers at each stage. Public benchmarks exist for property-level costs and onsite turnover. For almost every operating metric, there is none: the numbers that circulate are vendor claims or trade folklore, and we have not quoted them.

These are gaps in the public evidence as of October 2026. If you have data on any of them, we would like to hear from you.

Sources were read in October 2026; dates are publication or data dates. Last reviewed October 2026.