New research from Cloudera and Google/MIT shows that most stalled AI deployments trace back to data access and governance gaps, not model quality. For multifamily, where portfolio data is scattered across Yardi, RealPage, and MRI instances plus unstructured documents like leases and inspection reports, this is a familiar and costly problem. The operators seeing real ROI are the ones who fixed their data foundation before scaling AI, not after.
Every multifamily operator has a version of the same story. Leadership greenlights an AI pilot, maybe a leasing chatbot, maybe a predictive maintenance tool, maybe something meant to flag delinquency risk before it becomes a write-off. The demo looks sharp. Then it hits production and starts making recommendations that don’t match what the regional manager already knows, or it can’t see half the portfolio, or it takes so long to pull an answer that nobody bothers asking it twice. Six months later it’s quietly shelved.
New research out of Cloudera and a separate study from Google and MIT put numbers to a pattern most operators already feel in their gut. Across more than a thousand enterprise architects surveyed, the overwhelming majority had delayed or killed AI projects in the past year, often more than once, and the root cause traced back to data governance, access, and compliance gaps rather than the AI models themselves. A parallel survey of IT and product leaders found more than half had paused agentic AI deployments specifically to fix foundational data problems first.
That should sound familiar to anyone running technology for a multifamily portfolio.
In most industries, “data silos” means departments not talking to each other. In multifamily, it’s more literal. A single portfolio might run Yardi in one region, RealPage in another, and a legacy instance nobody’s gotten around to migrating off since the last acquisition. Rent rolls live in one system. Work orders live in another. Lease documents sit as scanned PDFs in a file share that predates the current VP of Operations. Delinquency data, reserve fund tracking, vendor contracts, inspection reports: all real, all valuable, and almost none of it structured in a way an AI agent can actually reason over.
The research frames this as a distinction between “data leaders” and “data laggards.” Organizations that expose the majority of their data to AI systems report consistently accurate, trustworthy outputs. Organizations that keep AI walled off from most of their data report the opposite: agents that hallucinate, miss context, or simply can’t complete the task. The gap isn’t subtle. It’s closer to a coin flip on whether the tool is useful at all.
For multifamily, this maps almost exactly onto the operational reality of a portfolio managed across disconnected platforms and property management companies. An AI tool that can only see one property management system’s data, in one region, formatted one way, isn’t going to catch the leasing velocity pattern that only shows up when you compare it against three other communities running a different platform. It’s not an intelligence gap. It’s a visibility gap.
The Google and MIT research also calls out unstructured or “dark” data as a major blind spot: information trapped in images, PDFs, and formats that don’t translate cleanly into a queryable dataset. Multifamily runs on exactly this kind of content. Lease abstracts. Inspection photos. Board meeting minutes. Vendor contracts. Move-in and move-out checklists. All of it carries operational signal, and almost none of it is sitting in a format an AI agent can natively use.
This is part of why so many early chatbot and automation pilots in property management underperform. The tool isn’t failing because the underlying model is weak. It’s failing because nobody built the layer underneath it that turns scattered, unstructured, multi-system data into something coherent enough to act on.
There’s a second finding worth sitting with. Nearly three-quarters of respondents said AI integration has made data governance more complex, not less. That tracks. Traditional access controls were built around the assumption that a human is the one requesting the data, at a reasonable pace, with judgment applied before anything gets used. Agentic AI breaks that assumption. Now it’s not one property manager pulling a rent roll. It’s an automated system potentially touching resident PII, financial records, and compliance-sensitive documents across an entire portfolio, continuously, without a person in the loop for every request.
For an industry already navigating fair housing compliance, financial audits, and board-level fiduciary obligations, that’s not a small wrinkle. It’s a reason governance has to be designed before AI gets deployed at scale, not retrofitted after something goes wrong.
The uncomfortable but useful takeaway from this research is that the operators seeing real ROI from AI aren’t the ones who moved fastest. They’re the ones who did the unglamorous work first: inventorying what data exists, understanding where it’s fragmented, cleaning it up, and building the context layer that lets an AI system actually understand what it’s looking at. Only then did they layer AI on top.
That’s a hard sequence to sell internally, because “get your data house in order” doesn’t generate the same excitement as “deploy an AI agent.” But the data backs it up. The portfolios treating AI readiness as a data problem first are the ones getting consistent, trustworthy results. The ones chasing the tool before fixing the foundation are the ones quietly walking back pilots six months in.
If there’s one question worth asking before the next AI initiative gets budget approval, it’s not “which tool should we buy.” It’s “can this tool actually see our data, and do we trust what it would find if it could.”