Top 10 AI Agents forProperty Managers in 2026

Every ranking of AI agents for property management promises the same thing: buy this and your team will handle more units with fewer people. The promise cannot be checked, because no independent benchmark of these products exists. We reviewed eighteen products on 23 August 2026 and found no academic study, no third-party head-to-head test, no audited leasing data and no trade-body evaluation of any commercial AI product sold into multifamily or single-family rental. Every containment rate, every conversion lift, every dollar of claimed delinquency reduction in this market is published by the seller about its own software.

That is a precise claim, not a lazy one. Some adjacent categories genuinely do have independent measurement. The nearest thing to it here measured drive-thru voice ordering rather than leasing: an Intouch Insight study in October 2025 put AI order accuracy at 83 percent against 87 percent for humans, reaching 95 percent only with staff intervention. It ranks no vendor in this market and settles nothing about apartments. What it does establish is that when someone finally measures one of these systems from outside, the number is lower than the marketing and the humans are still in the loop.

The strongest evidence in this entire category is not a benchmark at all. It is a public company writing to its investors under securities liability. AppFolio, Inc. (NASDAQ: APPF) filed its FY2025 Form 10-K on 5 February 2026. Under Item 1A, in its own risk factors, it wrote this:

Additionally, generative and agentic AI has the capacity to yield inaccurate or misleading results, promote discriminatory outcomes, or perpetuate unintended biases, which risks are exacerbated by the speed and scale of adoption and utilization of AI across our customer base.

— AppFolio, Inc., FY2025 Form 10-K, filed 5 February 2026

Read that against any vendor marketing page in this category, including AppFolio's own. The marketing says the AI is fair-housing aligned and monitors for bias. The filing says agentic AI can promote discriminatory outcomes and perpetuate unintended biases, and that the risk grows with scale. Both documents are from the same company, three months apart. The difference is the audience: one is written for buyers, the other for people who can sue over inaccuracy. This is not an admission of any specific failure and should not be read as one. It is the most cautious document in the market, and it is not the one the category quotes.

The operator's squeeze underneath all of this is real, which is why the pitch lands. The US Census Bureau's Quarterly Residential Vacancies and Homeownership release for the second quarter of 2026, CB26-116 of 28 July 2026, puts the national rental vacancy rate at 7.3 percent with a margin of error of plus or minus 0.2, against 46,827,000 renter-occupied units and 3,727,000 vacant units for rent. The regional spread matters more than the national figure for anyone running a real portfolio: South 9.5 percent, Midwest 6.9 percent, Northeast 5.9 percent, West 5.3 percent, principal cities 8.0 percent against suburbs 6.9 percent. Meanwhile the Bureau of Labor Statistics series for rent of primary residence (CUUR0000SEHA) moved from 435.489 in July 2025 to 447.963 in July 2026, a rise of 2.9 percent, while CPI-U all items rose 3.4 percent over the same window. Revenue per unit is growing more slowly than costs. That is the actual problem, and software can help with parts of it.

Three of the names that appear at the top of nearly every competing list are no longer independent products, and two of them are now owned by the same company. This is the fastest way to tell whether a 2026 ranking was actually re-checked in 2026 or copied from 2023 content, and it is the first thing an operator should look for.

• Colleen AI — acquired by Entrata, 20 June 2024: Entrata's press release is titled around ushering in a new era of autonomous property management, and states that through the acquisition of Colleen AI, Entrata is introducing ELI+. Terms were not disclosed. As of 23 August 2026, colleenai.com no longer resolves at all and colleen.ai redirects to Entrata's ELI product page. It is not a vendor you can buy from.
• Rent Dynamics — acquired by Entrata, 13 July 2023: Its own site now states the product is part of Entrata, and the RentPlus, CRM and contact-centre products are maintained by Entrata. Together with Colleen AI, that is two of the three dead names inside one acquirer.
• Knock — a RealPage product, not an independent CRM: The knockcrm.com brand still exists, which is why it keeps appearing in listicles as a competitor. But RealPage's own AI Leasing Agent page tells buyers to capture every interaction automatically in Knock CRM, and Knock's support links resolve to realpage.com. Rank RealPage; treat Knock as its CRM layer.

The practical consequence is not pedantic. An article ranking Colleen AI against Entrata ELI+ in 2026 is ranking a product against its own owner, and an operator who takes that shortlist into a procurement process will discover it in the first vendor call. It also changes the concentration picture: Entrata is the most acquisitive vendor in the category, and RealPage sits on 24 million rental units globally by its own account while AppFolio reports 9.4 million units under management in an audited SEC filing. Three of the ten products below are owned by companies that also own a large share of the systems of record they plug into.

We scored eight attributes, every one of which an operator can re-check without trusting us, and we refused to score performance at all. Everything below was checked on 23 August 2026. Where a cell says not publicly disclosed, that is a finding about the vendor's disclosure, not an estimate we declined to make.

The refusal in the last row is the point of this article, so it is worth naming what got refused. Here is a representative sample of the figures we collected from vendor sites in this category and did not print as fact: resolve up to 86 percent of inquiries without staff intervention; increase tour conversion rates by up to 30 percent; up to a 59 percent reduction in skips and eviction; 46 percent conversion for AI-handled prospects compared to 19 percent tour conversion; 2.5 times higher lead-to-visit conversations; a 10-second response time; 149 percent increase in property tours; 8.2 million hours added back to onsite teams; a 95 percent automation rate on workflows; and $1,680 to $2,592 per unit in maintenance savings.

Three things are wrong with that list, and they are wrong structurally rather than because any particular vendor behaved badly. First, the phrase up to makes a claim unfalsifiable: up to 86 percent is true if one property once hit it. That is a ceiling with no floor, not a performance figure. Second, the metrics are undefined. Conversion from what, to what, over what window? A figure comparing AI-handled prospects at 46 percent to tour conversion at 19 percent is comparing two different denominators. Third, none of them carries a sample, a control group, a time period or a method.

One more number gets refused, and it is the one every leasing-automation pitch is built on. “Resident turnover costs $3,000 to $5,000 a unit” appears across an enormous amount of published content. We chased it. It circulates simultaneously as $1,000 to $5,000, as $3,000 to $5,000, as a flat $3,900 and as one to three months' rent — and the last of those cannot be true at the same time as the others, because it makes the cost a function of the rent. When we went to the two trade bodies whose names are usually attached to it, the National Multifamily Housing Council's quick-facts pages returned a not-found page body while serving an HTTP 200 success code, and the National Apartment Association's income-and-expenses survey URL returned a 404. That soft-404 detail is worth a sentence on its own: a URL that returns success while serving a missing page will pass any automated citation checker, which is precisely how dead statistics stay in circulation for a decade.

What we print instead is the Census Bureau's rental vacancy rate, which publishes its own formula and its own margin of error, and the arithmetic you can run against your own rent roll in about four minutes. Your number will be true for your portfolio, which no industry average ever is.

Ten products, six verifiable columns, every cell either citable to the vendor's own page and a public filing or marked as not disclosed. Nothing in this table is a performance figure. Checked 23 August 2026.

Read the pricing column first, because it is the column with the least ambiguity. Two of the twelve products we assessed as rankable publish a number an operator can read without a sales call. Ten do not. The pattern holds across the compliance column too: one standalone vendor names a certification type on its own security page, two operate live trust centres whose contents are gated behind a request, and the rest publish nothing we could find. None of that is proof of anything about the underlying software. It is proof of what a buyer can establish before signing, which is a different and more useful thing.

The order below is verifiability, not quality, and the two are not the same thing. A product can be excellent and opaque, and several here probably are. What the order measures is how much of a vendor's claim an operator can confirm before money changes hands.

Considered and not ranked

Two products cleared our trading-status check but produced fewer verifiable cells than the ten above, so they sit here with their reasons stated rather than being quietly dropped. Funnel Leasing is a real and current vendor — Funnel Leasing, Inc., privacy notice last updated 12 January 2026 — with Yardi, RealPage and Entrata named as integrations and a renter-centric CRM plus an autonomous-as-described virtual leasing assistant. What it does not publish is any compliance artefact at all: no SOC 2, no ISO 27001, no DPA offer, no retention policy, no data-residency statement and no comprehensive sub-processor list, and its privacy notice concedes that no network or system is ever entirely secure. Its 46-percent-versus-19-percent conversion claim is the sharpest example in this market of two different denominators presented as one comparison. PERQ is trading and independent, with Yardi, RealPage, Entrata and Apartments.com named, but its public material describes no agent behaviour, no autonomy boundary and no approval model. On this article's own standard the honest classification is marketing automation with an AI label until the vendor documents otherwise.

Yardi requires a different kind of honesty. Yardi runs a large share of this market and is unquestionably a leader. But yardi.com, rentcafe.com and info.yardi.com returned HTTP 403 to every automated fetch on 23 August 2026, so no 2026 Yardi AI product name, autonomy description, integration list, trust page or fair-housing statement could be verified from Yardi's own site — and we will not write a product name from memory in a vertical where vendors rename products. What is verified: the legal entity is Yardi Systems, LLC, established from a 2026 federal filing by Yardi's own counsel noting the company was erroneously sued as Yardi Systems, Inc., with RentCafe, LLC a separate entity; and Yardi Breeze publishes real pricing at $1 per unit per month with a $100 monthly minimum, or Breeze Premier at $1 per unit per month with a $400 minimum, with RentCafe Chat IQ listed as a Premier add-on whose price is not stated. Naming the gap is better than filling it.

Three more were verified alive and left on the bench: Latchel (AI-first maintenance intake with disclosed human escalation and the widest named PMS list of any maintenance vendor, but no published pricing, no SOC 2 and no verifiable ownership), Snappt (application fraud detection, SOC 2 Type II stated on its own site, integrations named across Entrata, RealPage, Inhabit, ResMan and Yardi) and ResMan by Inhabit (not independent; plan names published, every figure gated, and whether AI features cost extra is not stated). Snappt earns a specific mention in the fair-housing section below, because of what it markets rather than what it builds.

An AI agent that talks to prospects, screens applicants or writes listing copy is doing the three things the Fair Housing Act regulates most tightly — and under the FHA the housing provider stays liable no matter how much of it was outsourced to software. That sentence is the whole section, and everything else here is HUD saying it in its own words.

HUD's Office of Fair Housing and Equal Opportunity published two guidance documents dated 29 April 2024, announced in press release HUD No. 24-098 on 2 May 2024. One addresses tenant screening, the other addresses advertising through digital platforms. Note the citation trap before you go looking: the live hud.gov URLs for both documents return 404 as of 23 August 2026, and only the archives.hud.gov copies resolve. The screening guidance states the operative principle directly:

Under principles of direct liability, housing providers are responsible for ensuring their rental decisions comply with the Fair Housing Act even if they have largely outsourced the task of screening applicants to a tenant screening company. The screening company may be responsible too, but housing providers retain authority over screening practices and decisions at their properties.

— HUD FHEO, Guidance on Application of the Fair Housing Act to the Screening of Applicants for Rental Housing, 29 April 2024

HUD goes further and explains why the vendor is also exposed, quoting the statute's structure: by drafting the Act in the passive voice, Congress banned an outcome while not saying who the actor is, so any entity that plays a substantial role in a housing decision can be liable even if it is not the ultimate or sole decisionmaker — including those with input into what variables are considered or what standards should be used. It also records that a person is vicariously liable for a discriminatory housing practice by that person's agent or employee regardless of whether the person knew or should have known. If you are shopping for a vendor to absorb your fair-housing risk, HUD has already explained that no such vendor exists.

HUD also anticipated the exact marketing move this category makes. Its screening guidance notes that some tenant screening companies advertise that automated screenings are less likely to be discriminatory and that their services can help housing providers comply with the law, footnoting Conn. Fair Hous. Ctr. v. CoreLogic Rental Prop. Sols., LLC, 478 F. Supp. 3d 259, 275 (D. Conn. 2020). Set that against a live 2026 vendor page: Snappt markets itself as Fair Housing Compliant with an independent applicant-agnostic view that mitigates any fair housing or discrimination issues. The word “any” is doing impossible work there. HUD naming the marketing claim, and a current vendor making it, is a complete argument in two citations.

Disparate impact: get the tense right

HUD has not eliminated disparate impact, and writing that it has would be seriously wrong. HUD published a proposed rule on 14 January 2026 (FR-6540-P-01, RIN 2529-AB09) stating that it is proposing to remove its discriminatory effects regulations and leave to courts questions related to interpretations of disparate impact liability under the Fair Housing Act. Comments closed 13 February 2026. It then published a supplemental notice of proposed rulemaking on 10 August 2026 which reopens the January comment period, with comments due 9 October 2026.

Four things follow, and an operator should hold all four. First, both documents are proposals. No final rule has issued, and 24 C.F.R. § 100.500 is in force today. Second, the three-step burden-shifting framework still governs: a plaintiff shows the practice causes or predictably will cause a disparate impact; the burden shifts to the defendant to show the practice is necessary to achieve a substantial, legitimate, nondiscriminatory interest that may not be hypothetical or speculative; then it shifts back to show a less discriminatory alternative. HUD adds a line built for algorithmic screening: if a screening policy is overbroad, meaning it screens out unproblematic applicants, a more targeted policy could be a less discriminatory alternative. Third, even a finalised repeal would not end the exposure — HUD's own abstract says the question would go to the courts, and Inclusive Communities, 576 U.S. 519 (2015), is a Supreme Court construction of the statute that an agency cannot repeal by rulemaking. Fourth, there is precedent for how these attempts go: the 2020 version of the discriminatory effects rule was enjoined before it took effect, so the 2013 framework has been in effect without interruption since it was issued.

The practical advice for an operator is therefore identical in either world, and that is the honest close. Private plaintiffs and state attorneys general bring these cases. SafeRent was a private class action. Many states have their own fair-housing statutes with their own effects tests. Nothing about a federal rulemaking changes what your leasing agent is permitted to say.

Listing copy: where intent stops mattering

This is the most under-covered exposure in the category and the easiest one for an operator to walk into without noticing. 42 U.S.C. § 3604(c) prohibits making, printing or publishing any housing-related notice, statement or advertisement indicating a preference, limitation or discrimination based on a protected characteristic. 24 C.F.R. § 100.75(c)(1) reaches descriptive copy directly, prohibiting words, phrases, photographs, illustrations, symbols or forms conveying that dwellings are available or not available to a particular group. And HUD, at footnote 19 of the advertising guidance, records the decisive point: courts consistently interpret this to mean a defendant can violate § 3604(c) if the notice indicates discrimination to an ordinary reader or ordinary listener, regardless of whether the defendant intended to discriminate.

A generative model has no intent at all, and § 3604(c) does not require any. Copy praising a family-friendly building, a quiet mature community, a great Christian neighborhood, a perfect starter home for a young couple, or a neighbourhood described by the demographics of who lives there, is judged by what an ordinary reader takes it to mean. The model's lack of intent is not a defence and neither is the vendor's. HUD's advertising guidance also warns that ad delivery can skew without the advertiser's direction or knowledge and can even frustrate an advertiser's intention that an ad be distributed more broadly, and it lists steering home-seekers to particular neighborhoods among the ways targeting violates the Act. Its definition of advertiser expressly covers rental housing and property management services, so this is not somebody else's problem.

Screening is the highest-liability workflow in this vertical because two separate legal regimes land on one process, and satisfying one does not satisfy the other. HUD says so expressly in footnote 1 of its screening guidance: some tenant screening practices are governed by other federal laws, such as the Fair Credit Reporting Act, 15 U.S.C. §§ 1681–1681x, the obligations under which are not covered by that guidance. Fair Housing Act compliance is not FCRA compliance.

The operator-facing FCRA obligation, stated at the level of generality the primary sources support: if a denial, a higher deposit, a co-signer requirement or any other adverse action is taken based in whole or in part on a consumer report, the applicant is owed an adverse action notice identifying the reporting agency, stating that the agency did not make the decision, and informing the applicant of the right to a free copy of the report and the right to dispute its accuracy. The FTC publishes guidance for both sides of this — one resource for landlords using consumer reports and one for tenant background screening companies — and the CFPB publishes FCRA and Regulation V compliance resources. HUD points readers to all of them.

Now the AI-specific failure, and it is HUD's own observation rather than ours. These technologies, HUD writes, can lead to a less transparent process by obscuring the precise reasons for a denial from the housing provider and the applicant. An adverse action notice must convey reasons. A model that cannot explain its own denial cannot support the notice the law requires. That is not a philosophical objection to machine learning; it is a procurement question with a yes-or-no answer, and it is the single most useful question to ask a screening vendor: show me the reason codes your model produces, in the words my adverse action notice will use.

Pair that with the gap recorded above. RealPage's AI Screening page describes predictive risk scoring on more than 30 million lease outcome records, assessing willingness to pay rather than credit score alone, and carries no FCRA language, no consumer-reporting-agency statement, no adverse-action reference, no dispute-rights reference and no fair-housing language at all. We are not alleging non-compliance; the page is marketing, not a compliance filing, and obligations can live elsewhere. We are reporting a verified absence in the highest-liability workflow in the vertical, and noting that a buyer reading only that page would come away with no idea that any of these obligations exist.

What Louis v. SafeRent actually decided

It settled. There was no merits adjudication, no trial, no finding of liability and no admission of wrongdoing. Louis v. SafeRent Solutions, LLC, No. 1:22-cv-10800-AK (D. Mass.), before Judge Angel Kelley, was filed in May 2022 by two Black rental applicants using housing vouchers, against SafeRent Solutions, LLC (formerly CoreLogic Rental Property Solutions) and a property management company. The claims were Fair Housing Act disparate impact plus Massachusetts anti-discrimination law. It settled for $2.275 million.

The allegation, in the Justice Department's own words on its case page, is that the plaintiffs were denied due to their SafeRent Score, a score derived from an algorithm-based screening software, and that the scores result in an unlawful disparate impact against Black and Hispanic applicants because the underlying algorithm relies on factors that disproportionately disadvantage them, such as credit history and non-tenancy related debts, while failing to consider one highly relevant factor: that the use of HUD-funded housing vouchers makes such tenants more likely to pay their rent. The United States filed a Statement of Interest on 9 January 2023 clarifying that the FHA applies to companies providing residential screening services.

The only merits-adjacent ruling came on 26 July 2023, when the court denied the motions to dismiss the FHA claims, holding that SafeRent was subject to the FHA and that the plaintiffs had alleged a plausible claim for disparate impact discrimination against both defendants. HUD's own 2024 guidance cites that decision twice, which is the cleanest evidence that it mattered. A denial of a motion to dismiss decides only that a claim was plausibly pleaded. The correct description is: the first algorithmic tenant-screening case to reach a class settlement, and the one where a court confirmed a screening vendor can be sued under the Fair Housing Act at all. That is exposure, not proven discrimination — and it is more than enough to change how you configure a screening workflow.

One structural detail is worth carrying away. The FHA does not cover source of income. The voucher-holder theory reached the plaintiffs through Massachusetts state law running alongside the federal race-based disparate-impact claim. Source-of-income protection is state and local law, it exists in a large and growing number of jurisdictions, and an agent configured once for a national portfolio will be wrong somewhere. Check your own jurisdictions rather than trusting a vendor default. And note HUD's under-used line while you do: screening practices prohibited by any applicable law will not be considered necessary to achieve a substantial, legitimate, nondiscriminatory interest under the Fair Housing Act — so violating a local screening ordinance also forfeits your federal business-necessity defence.

Keep this separate from leasing and screening AI throughout. The Justice Department's case is about revenue management software, and nothing in it concerns AI leasing agents. Conflating them is a factual error and also unfair to the vendors. But if you run units in New York, San Francisco or Berkeley, the rent-setting rules bind your portfolio today regardless of what your leasing agent does.

The federal case, with the status precision that most coverage gets wrong. United States and Plaintiff States v. RealPage, Inc., No. 1:24-cv-00710-WLO-JLW (M.D.N.C.), was filed 23 August 2024 against RealPage alone and amended 7 January 2025 to add six landlord defendants, alleging Sherman Act § 1 information-sharing and price alignment, and § 2 monopolisation against RealPage. As of the Antitrust Division's case page last updated 6 July 2026, exactly one final judgment has been entered: Greystar Management Services, LLC, entered 2 March 2026. RealPage's own proposed final judgment, filed 24 November 2025, had not been entered. Neither had LivCor's (filed 23 December 2025), Willow Bridge's (6 July 2026) or Cortland's (7 January 2025).

What the proposed RealPage judgment would require is a better lede for an operator than the fact that a case exists, because it is the first set of concrete design constraints ever imposed on a US pricing model. Per DOJ's press release of 24 November 2025, if approved it would require RealPage to cease having its software use competitors' nonpublic, competitively sensitive information to determine rental prices in runtime operation; to cease using active lease data for model training, limiting training to historic data aged at least 12 months; to avoid models determining geographic effects narrower than state level; to remove or redesign features that limited price decreases or aligned pricing between competing users; to cease market surveys collecting competitively sensitive information; to accept a court-appointed monitor; and to cooperate in the United States' suit against management companies. Assistant Attorney General Abigail Slater said that competing companies must make independent pricing decisions, and that with the rise of algorithmic and artificial intelligence tools the Division will remain at the forefront of vigorous antitrust enforcement.

New York is the big one, and it is in force. N.Y. General Business Law § 340-b, added by S.7882 of the 2025 session, was signed 16 October 2025 and took effect 15 December 2025. It makes it unlawful to operate, license or use software recommending rental prices, lease renewal terms, ideal occupancy levels or other lease terms to a residential owner or manager where it does so by analysing or processing data from two or more residential rental owners or managers, and separately makes it unlawful for owners and managers to set or adjust rents on such recommendations knowingly or with reckless disregard. Penalties reach up to $1 million per unlawful act under GBL § 341, plus criminal exposure. RealPage's First Amendment challenge, RealPage, Inc. v. James, No. 1:25-cv-09847-VEC (S.D.N.Y.) before Judge Valerie E. Caproni, was filed 26 November 2025; its preliminary-injunction motion and the Attorney General's motion to dismiss were both fully briefed by 27 February 2026 and remain pending and undecided, with supplemental-authority letters filed 10 and 13 August 2026. The law has not been enjoined. No court has blocked it, and no court has upheld it. The correct word is pending.

San Francisco was first, and it is the one that stacks. Ordinance No. 224-24, finally passed 3 September 2024 and codified at San Francisco Administrative Code § 37.10C, prohibits both selling such a device to San Francisco landlords and a landlord using one, and provides that each separate month a violation exists or continues, and each separate residential dwelling unit for which the landlord used the device, constitutes a separate and distinct violation. Penalties run up to $1,000 per violation, and a tenant has a private right of action for injunctive relief, damages and penalties with mandatory attorney's fees to a prevailing tenant. Per-unit, per-month stacking plus a private right of action is what makes this operationally serious for a mid-size operator with San Francisco exposure. The ordinance defines its target narrowly around non-public competitor data and expressly disclaims any intention to prevent software that helps landlords manage units generally or through public data, or to regulate the amount of rent charged.

Berkeley shows where this is heading. Ordinance No. 7,956-N.S. banned software recommendations on lease terms; RealPage sued in RealPage, Inc. v. City of Berkeley, No. 25-cv-3004 (N.D. Cal.), filed 2 April 2025; Berkeley then agreed to amend, producing Ordinance No. 7,992-N.S., which — in RealPage's own characterisation in its later briefing, and it is a litigation position rather than a neutral description — regulates how software works and analyzes data to ensure it cannot facilitate anticompetitive conduct, without categorically banning speech. That is the same direction as the DOJ remedy: away from banning the recommendation and toward constraining the data and the model. An operator's compliance question in 2027 will not be whether rent software is legal here, but what data this model touches.

California is narrower than the headlines. AB 325 (Aguiar-Curry), Chapter 338, Statutes of 2025, chaptered 6 October 2025, adds Business and Professions Code §§ 16729 and 16756.1. It makes it unlawful to use or distribute a common pricing algorithm as part of a contract, combination or conspiracy to restrain trade, or to use one if the person coerces another to adopt a recommended price or term. That is not a flat ban. The half that matters more to defendants is procedural: it lowers the Cartwright Act pleading standard so a complaint need only allege facts making a conspiracy plausible, and is not required to allege facts tending to exclude the possibility of independent action. And Colorado's rent-algorithm bill was vetoed on 30 May 2025 — do not list Colorado as a ban.

Beyond those four, we did not verify. Philadelphia, Minneapolis, San Diego, San Jose, Jersey City, Providence and Hoboken appear in trackers; municipal code libraries blocked automated access from our environment, so we will not name them as in force. If you operate there, open the municipal code yourself. Naming an unverified city as having an in-force ban is exactly the error this article exists to refuse.

An AI voice agent is an artificial voice under the TCPA, and the rules that apply to it did not get easier in 2025. FCC Declaratory Ruling 24-17, CG Docket No. 23-362, adopted 2 February 2024 and released 8 February 2024, grounds this in 47 U.S.C. § 227(b)(1)(A) and (b)(1)(B) and states, at paragraph 6, that the TCPA does not allow for any carve out of technologies that purport to provide the equivalent of a live agent, thus preventing unscrupulous businesses from attempting to exploit any perceived ambiguity in the rules. No commenter opposed those conclusions.

The trap, stated because it is the most damaging error available in this whole subject. Insurance Marketing Coalition v. FCC (11th Cir., 24 January 2025) vacated only Part III.D of the FCC's 2023 Order — the one-to-one consent provision. The court's own footnote states that the 2012 Order is not at issue. Prior express written consent for marketing calls and texts survives untouched. Writing that the TCPA got easier for AI callers inverts the law, and it is advice an operator could act on and be sued for. The correct summary: one narrow consent provision was vacated; the written-consent regime and the artificial-voice classification both stand, and an AI leasing voice agent sits squarely inside them. Status precision matters here — vacated (Part III.D) is not the same as the rest of the order, which is not the same as the 2012 Order.

This bites harder in leasing than almost anywhere else, because the core loop of an AI leasing agent is outbound follow-up to inbound leads — often at odd hours, often by SMS, often persisting for weeks. That is marketing to a consumer. Prior express written consent, caller identification, disclosure of the party responsible for the call, and a working opt-out are not optional. And there is a structural problem specific to this vertical: leads sourced from listing sites arrive with consent obtained by a third party. The operator carries the risk, not the listing site. Several states also impose their own consent rules with private rights of action; unless you have read the specific statute, treat that generically and check your own jurisdictions.

This is an illustrative worked scenario, not a client outcome. No client is described here, and no result is claimed. Every input is a published figure from a vendor pricing page or a federal statistical release, and the arithmetic is shown so you can redo it with your own numbers.

Consider an operator running 1,400 units across four markets — a realistic mid-size portfolio, above the point where every vendor minimum stops binding and below the point where enterprise pricing gets individually negotiated. The question is not what AI would save them. Nobody can compute that from public data, which is the finding. The question is: how much of their AI spend can they price before signing anything?

Maintenance, priced exactly. Property Meld Ops at $2.00 per unit per month is 1,400 × $2.00 = $2,800 a month, or $33,600 a year. The $160 monthly minimum is irrelevant at this size. Adding MAX On-Call for after-hours triage at $1.50 per unit per month adds 1,400 × $1.50 = $2,100 a month, or $25,200 a year. Total: $4,900 a month, $58,800 a year. Choosing Core at $1.60 instead of Ops would be $26,880 a year before the add-on. That entire calculation takes about ninety seconds and requires no sales call.

Touring and lead nurture, priced with one stated assumption. Rently charges per active listing per month. To estimate active listings we need a vacancy assumption, and here is where the article's own rule applies: use your own rent roll, not a national average. For the sake of showing the arithmetic we will substitute the Census Bureau's Q2 2026 national rental vacancy rate of 7.3 percent, and flag plainly that it is a placeholder and that the regional figures range from 5.3 percent in the West to 9.5 percent in the South. At 7.3 percent, 1,400 units implies roughly 102 units listed at any given time. At one credit per active listing per month that is about 1,224 credits a year. Premium at $37 is $45,288 a year; Standard at $28 is $34,272. The difference — 1,224 × $9 = $11,016 a year — is the only separately priced AI layer we found anywhere in this market. Every other vendor bundles it into an undisclosed platform fee.

The system of record, where the sum stops. Yardi Breeze Premier at $1 per unit per month is 1,400 × $1 = $1,400 a month, or $16,800 a year, comfortably above the $400 monthly minimum. But the price of the RentCafe Chat IQ AI add-on is not stated, so the operator cannot complete the line. That is the pattern in miniature: the PMS is priced, the AI on top of it is not.

The total, and the hole in it. Published spend across the two vendors that publish a price: roughly $104,088 a year — $58,800 for maintenance plus $45,288 for touring and nurture. Products for which no annual figure can be computed from any public source: AppFolio Realm-X, Entrata ELI+, RealPage Lumina, EliseAI, Travtus, Zuma, ResiDesk, Hyly.AI, Funnel and PERQ. Ten of twelve. An operator can walk into procurement with an exact number for two of them and nothing at all for the other ten.

The break-even you can actually run. Divide $104,088 by your own average monthly rent and you have the number of unit-months of vacancy the two products must remove to pay for themselves before any labour saving. Use your number, not ours — but to show the shape: at a placeholder average rent of $1,500 a month, that is about 69 unit-months a year, against roughly 1,224 listing-months in the same period. A little under 6 percent. That is a defensible target to test in a pilot at one property. It is not a promise, and no vendor number was used to produce it. Notice what this frame does: it prices the decision on cost, verifiability and exit rather than on a return nobody can substantiate. If you are weighing this against building the workflow yourself, our build-versus-buy analysis for AI agents works through the same comparison with the internal-cost side filled in.

The failures that matter in this vertical are not model failures. They are consent failures, copy failures and permission failures, and every one of them has a detection signal you can instrument before launch.

The fourth row deserves expansion because the category consistently understates it. Prompt injection is unsolved. An agent that reads resident messages, maintenance photos, uploaded income documents, vendor invoices or emails is reading untrusted input written by people who may have an interest in what the agent does next. There is no filter that reliably distinguishes instructions embedded in content from content itself, and anyone selling you one is selling a mitigation as a cure. The correct posture is blast-radius reduction: give the agent the narrowest write credential that lets it do its job, scope that credential to specific object types, log every tool call, and make credential revocation — not code deployment — the rollback lever, because a paused deployment can be restarted by a queued job while a revoked credential cannot act. We work through the current threat taxonomy in our guide to prompt injection and the OWASP LLM Top 10.

Write this table into your standard operating procedure before you write a single prompt, because it is the document that will be read if anything goes wrong. The columns are deliberately blunt: what the agent may do without anyone watching, what requires a named human before it takes effect, and why the boundary sits where it does.

Two clarifications that vendors blur. First, an interrupt is not an approval. A system that pauses and alerts a team member when a prospect responds has summoned a human by event; it has not asked anyone to authorise the outbound message that already went. Both designs are legitimate; only one of them puts a person between the model and the consumer, and you should know which one you bought. Second, disclosed hybrid is a feature, not a defect. Two vendors here publish that human teams back their AI. Given what the SEC order in Presto actually punished — concealment of human involvement, not human involvement — a disclosed human layer should read to an operator as a credibility signal.

These are Frenchy Digital's 2026 scoping bands for senior-led delivery, not an industry benchmark. Read them the way you should read every other number on this page: as one seller describing its own pricing. The difference is that we are telling you that.

Senior-led delivery runs $150 to $225 per hour, ongoing retainers run $2,500 to $9,500 per month, and every engagement carries a 30-day post-launch warranty. We return a fixed-price phased proposal within 5 business days of a discovery call. Full source-code and IP ownership transfers to you. We are a senior-led Black-owned agency in Los Angeles, reachable at +1 (424) 272-5601 or at calendly.com/frenchydigital/discovery-call.

For most operators between 200 and 5,000 units, the honest recommendation is not a custom build at all. Buy the maintenance triage, because it is priced, low-risk and the workflow is bounded. Buy or build the resident messaging layer depending on how many PMS platforms you run. Then spend the discovery budget on the two things no vendor will do for you: mapping your fair-housing exposure across prospect-facing copy and screening, and rebuilding your consent record so outbound follow-up is defensible. Those two artefacts survive every vendor change, and they are the ones a plaintiff's lawyer will ask for.

Nine things that should slow a procurement down, drawn from what we actually found on vendor sites in this category rather than from a generic checklist.

• A fair-housing badge with no method behind it: Compliance claims are more dangerous than performance claims, because an operator may rely on them. Three vendors here assert fair-housing safety and none publishes a bias-testing methodology, an audit or a disparate-impact analysis. Ask for the method in writing. No vendor can sell you fair-housing compliance, because under the FHA the liability is yours.
• The words up to in front of a percentage: Up to 86 percent is true if one property once hit it. It is a ceiling with no floor. Ask for the floor, the median, the sample size and the time period, and watch what happens.
• A certification cited without a type, an auditor or a report date: SOC 2 is not SOC 2 Type II. Alignment is not certification. And check whose certificate it is: elsewhere in this cluster we found a vendor citing ISO 27001 and SOC 2 that belonged to its cloud hosting provider, not to the vendor's own legal entity.
• A trust centre whose contents are gated behind a request: Two vendors here operate live trust centres that return nothing readable to a prospective buyer. That is not evidence of a problem; it is evidence you cannot check before signing. Make the report a condition of the pilot.
• Integration claims with no PMS named and no API reference: Integrates with your existing systems is not an integration list. Ask which named PMS, at which API tier, for read or for write. AppFolio's own plan table shows why the tier question matters: read-only on Plus, read/write only on Max.
• A screening product page with no FCRA or adverse-action language: This is a verified absence in this category, not a hypothetical. If the marketing page for a scoring product says nothing about consumer reports, adverse action or dispute rights, ask where those obligations are handled and get the answer in the contract.
• An agentic system with no described approval step anywhere: Absence of a documented boundary is itself information. Entrata publishes partial-to-full automation as a choice; several competitors describe autonomous sending and never mention a human at all.
• Any claim built on a resident-turnover cost benchmark: It circulates as $1,000 to $5,000, as $3,000 to $5,000, as $3,900 and as one to three months' rent simultaneously. The two trade bodies usually credited served us a soft 404 and a hard 404 respectively. Do not let a business case rest on it.
• A product name that stopped existing: Colleen AI, Rent Dynamics and Knock as an independent CRM. If a vendor, consultant or listicle pitches you one of these in 2026, everything else they told you needs re-checking too.

Limitations: what we could not verify

Stating this plainly is the price of the rest of the article. Eleven things we could not establish, listed so you can weigh them:

• Yardi's 2026 AI product line: yardi.com, rentcafe.com and info.yardi.com returned HTTP 403 to every automated fetch. Yardi Breeze pricing was verified; the enterprise AI product names, autonomy model and integration list were not, and we refuse to write them from memory.
• Whether HUD has withdrawn or reaffirmed its April 2024 AI guidance: The live hud.gov URLs for both documents now 404 and only the archive copies resolve. That is suggestive and proves nothing. We assert neither withdrawal nor reaffirmation.
• The duration and exact wording of the SafeRent injunctive relief: SafeRent reportedly agreed to stop using a SafeRent Score for Massachusetts applicants using housing vouchers, with a reported term of at least five years. The settlement administrator's site and the litigation clearinghouse both blocked access, so we state the relief without a duration.
• Whether RealPage's own Final Judgment has been entered since 6 July 2026: As of that DOJ case-page update it had not been. If you are relying on this, check the case page rather than this article.
• Rent-algorithm ordinances outside four verified jurisdictions: Only New York State, San Francisco, Berkeley and California were verified. Municipal code libraries blocked automated access, so Philadelphia, Minneapolis, San Diego, San Jose, Jersey City, Providence and Hoboken are named here as unverified rather than as in force.
• Any eviction statistic: The most-cited source blocked access from our environment, and its coverage is not a nationally representative sample while its filing counts include serial filings against the same household. A number whose denominator we cannot describe does not go in this article.
• SOC 2 status for EliseAI and Hyly.AI: Both operate live trust centres whose contents are gated. Neither the presence nor the absence of certification should be asserted, and we assert neither.
• Legal entity and ownership for most of the roster: Zuma, ResiDesk, PERQ, Hyly.AI, Travtus, Latchel, Snappt, Property Meld and Rently publish no legal entity or ownership information, and none was found in public record in this review. Funnel Leasing's entity name comes from its own privacy notice.
• The subject matter of two 2026 federal actions naming Yardi Systems, LLC: Both were verified to exist on the docket. Neither was read. We do not characterise them, and specifically we do not call them antitrust cases.
• Recent trade coverage: This research was built from direct fetches of primary documents — Federal Register entries, court dockets, SEC filings, HUD guidance, FCC orders, Census and BLS releases and vendor pages — rather than from search summaries. That makes the legal and ownership findings unusually strong and leaves a blind spot around the most recent quarter of private-company news.
• Any independent evaluation of any product in this table: None exists that we could find, and we looked. If someone shows you one, read who paid for it before you read the result.

None of this argues against deploying AI in a rental portfolio. It argues for deploying it in the order of consequence — internal copilots first, resident messaging second, prospect outbound third against a consent record you can produce, listing copy fourth with a named approver, and screening and renewal pricing last or never — and for buying on cost, verifiability and exit rather than on a return nobody can substantiate. The operators who get value from this work are the ones who found out what they were permitted to do before they found out what the model could do.

And the boundary holds throughout. An agent reads, classifies, routes, schedules, drafts and proposes. A named human commits anything that publishes, anything that denies an applicant, anything that sets a rent, and anything that contacts a consumer without a consent record you can produce on demand. Housing providers are responsible for ensuring their rental decisions comply with the Fair Housing Act even where the task has largely been outsourced — HUD has already said so in writing, and no vendor contract changes it.