AI is moving quickly, and real estate operators are sprinting to figure out how to put it to work. But as fast as new point solutions are being released, foundational models themselves are improving even faster. This is presenting a real estate operators with a new temptation: why not just build it ourselves?
After all, models like Claude, Gemini, and GPT can now draft lease summaries, triage maintenance requests, parse vendor contracts, and generate custom reporting dashboards with minimal engineering overhead. When Yardi announced its MCP connector to Claude last September, it removed one of the last meaningful technical barriers for large operators — suddenly data that once required engineering involvement to query is accessible through a conversation.
But “can do” and “should do” are different questions. The same capabilities that make foundational models exciting also make them dangerous in the wrong context. Multifamily operations touch a dense layer of regulation where a model that gets something wrong doesn’t just create a bad outcome, but real liability.
Today’s report walks through three use cases that illustrate where each approach wins: tenant screening, vendor management and contract abstraction, and custom reporting dashboards.
The Case for Point Solutions: Tenant Screening
Tenant screening is where most operators should stop before spinning up a foundation model.
The legal exposure here is real and well-documented. Fair housing law, the Fair Credit Reporting Act, and a patchwork of state-level regulations tightly govern what questions can be asked, what data can be used, and how adverse action decisions must be communicated. A screening tool that surfaces the wrong data point doesn’t just create operational headaches, it creates fair housing complaints, class action exposure, and reputational risk that dwarfs whatever you might save by not buying a purpose-built tool.
This is particularly true in income verification and fraud detection, where the margin for error is razor-thin and the regulatory landscape keeps shifting. Tools like Snappt, Vero, and built-in PMS screening tools have built their entire products around detecting document fraud and managing compliance obligations simultaneously. They’ve absorbed the liability of knowing what the rules are across jurisdictions and updating as those rules change. That’s not something you can replicate with a well-prompted Claude instance, regardless of how capable the underlying model has become.
The failure mode here is asymmetric. An operator might process 10,000 applicants without incident. But when something goes wrong, the exposure is severe, and “we built it ourselves with AI” is not a defense that plays well in a courtroom.
Verdict: Buy. The guardrails aren’t a product feature; they *are* the product.
Vendor Management and Contract Abstraction
Vendor contracts sit in an interesting middle ground. There’s legal exposure — terms, indemnification clauses, insurance requirements — but the stakes on any individual error are lower and more recoverable than a fair housing violation.
A foundational model can do meaningful work here. Extracting key terms from a stack of service agreements, flagging renewal dates, summarizing scope-of-work language across vendors, identifying contracts missing required insurance provisions — this is well within what current models handle accurately. For operators who have in-house or outside counsel reviewing outputs, the risk of an AI error is manageable.
But there’s a dimension to this use case that purely DIY implementations miss: market context. Purpose-built vendor management tools like Revyse don’t just abstract contract terms — they bring in comparative market data that an individual operator simply doesn’t have on their own. Knowing that your HVAC maintenance contract is priced 22% above market for your region, or that your landscaping vendor’s scope-of-work is unusually narrow relative to comparable properties, turns a contract review into a procurement optimization. A foundational model working only on your internal data can’t give you that.
The calculus also depends on operational maturity. If a property manager will treat AI-extracted contract summaries as final without review, purpose-built tools with audit trails and approval workflows earn their cost. If there’s a legal or operations team that treats AI output as a first pass, a custom implementation can work well. The technology isn’t the constraint — the process discipline around it is.
Verdict: Depends on your team and your goals. If comparative market intelligence matters to your procurement process, a point solution like Revyse justifies itself quickly. If you have strong review processes and primarily need contract organization, a foundation model handles it.
The DIY Case: Just-in-Time Dashboards
This is where foundation models have quietly become the most practical tool available — and where a recent product announcement changed the math considerably for large operators on Yardi.
In September 2025, Yardi announced Virtuoso Connectors at its annual YASC conference: a way to securely connect real-time Yardi data and tools with Claude through the Model Context Protocol (MCP), enabling property and asset management professionals to have natural language conversations with AI informed by their Yardi data. For an industry where a substantial share of large multifamily operators run on Yardi, this matters. The technical integration that previously required developer involvement now has a supported, enterprise-grade path.
Most operators’ instinct is that this is still a technical undertaking: something for the IT team, not the VP of Asset Management. And while that may have been true six months ago, it’s increasingly incorrect today.
With Virtuoso Connectors configured, an asset manager can open Claude and ask a question in plain English: “Show me rent-to-income ratios for lease renewals across my Class B portfolio, broken out by unit type, compared to in-place rents from 18 months ago.” Claude queries the Yardi data directly, structures the analysis, and returns a formatted response. Yardi’s framing for this is “Bring Your Own AI Assistant” workflows, with direct access to financial modeling, predictive maintenance, market analysis, and portfolio insights without manual data extraction.
For operators not on Yardi, it’s only a matter of time. Appfolio has built their own robust suite of AI-powered asset management tools, making this scenario realistic for operators on that platform as well.
The failure mode in this scenario is key. If a dashboard calculation is wrong, an operator catches it, asks a follow-up question, and gets a corrected version. Nothing irreversible happens. Compare that to screening or lease execution, where a bad AI output that reaches an applicant or a signed document creates real liability. Custom reporting sits entirely on the right side of that line — high variability in what you need from week to week, low stakes when the output is off, and zero point solution that can anticipate every ad hoc question your team will have.
Verdict: Build. The Yardi MCP connector removes the last meaningful technical barrier for large operators, and this is exactly the kind of high-value, low-risk use case that foundation models are built for.
On Operator Scale
Most of the analysis above is framed around operators large enough to negotiate enterprise contracts and deploy dedicated IT resources. But the build-vs-buy calculus actually tilts further toward DIY for smaller firms, and the reasons are worth spelling out.
Start with cost. Many point solutions come with minimums and per-property charges that make sense at scale but become punishing for a 200-unit portfolio. A tenant screening platform, a contract management tool, a BI dashboard, and a maintenance triage system can easily run five or figures annually when you add them all up. A foundation model API, on the other hand, costs pennies (or less) per query. Even factoring in the time to configure it, the total cost of ownership for low-risk use cases is dramatically lower.
Then there’s the simplicity advantage that smaller operators tend to undervalue. The reason large operators need purpose-built tools with elaborate permissioning, audit trails, and role-based access controls is that they have hundreds of people touching the system. A regional operator with a lean team doesn’t carry that same organizational risk. When the person building the AI workflow is the same person reviewing the output, the guardrail requirements shrink considerably. You don’t need an enterprise approval chain when your approval chain is a conversation across the office.
The rapid improvement in foundational models also matters disproportionately for smaller firms. Eighteen months ago, getting useful output from an AI tool required meaningful prompt engineering, custom integrations, and often a developer on call. Today, an operator with no technical background can have a productive conversation with Claude about their lease portfolio or maintenance backlog in plain English. Each new model generation makes the barrier to entry lower. That’s a bigger deal for the firm that doesn’t have an IT department than for the one that does.
None of this changes the overarching framework. Tenant screening and other high-liability use cases still demand purpose-built compliance infrastructure regardless of portfolio size. But for everything on the “build” side of the line, smaller operators may actually be better positioned to move fast with foundation models than their larger competitors, precisely because they have less organizational complexity to work around. The playing field is more level than it has ever been.
The Framework
When evaluating whether a foundation model or a point solution is right for a given use case, three questions do most of the work:
1. What’s the failure mode, and is it recoverable? Screening a tenant using flawed criteria: not recoverable. Generating a dashboard with a calculation error: recoverable. The higher the stakes on a single bad output, the more you want purpose-built compliance and audit infrastructure around it.
2. Is the output going to humans or to systems — and with how much review in between? If AI-generated output goes directly to an applicant, a vendor, or a legal document without human review, error tolerance needs to be near zero. If it goes to a person who will evaluate it before acting, normal error rates are acceptable.
3. Does the use case require domain-specific compliance knowledge or market data you don’t have? Fair housing, FCRA, state screening regulations, market pricing benchmarks — these aren’t things a general-purpose model will get right reliably without significant, jurisdiction-specific fine-tuning. Point solutions in these categories carry that burden so operators don’t have to.
The operators who will get this right aren’t the ones who build everything or the ones who buy everything. They’re the ones who correctly identify which use cases carry real legal or operational exposure if the AI gets it wrong — and buy guardrails there — while moving fast with custom implementations everywhere else. The Yardi MCP announcement is a signal that the “everywhere else” category just got a lot larger.
– Brad Hargreaves
