Smarter Systems, Stronger Communities

In recent years, Artificial Intelligence (AI) technology has seen significant progress and increased investment. By 2025, Meta, Amazon, Alphabet, and Google are projected to spend around $300 billion on AI research and development, according to CNBC.

In the affordable housing sector, AI usage falls mainly under the umbrella of two specific technologies: Large Language Models and Predictive Analytics. But as AI-based platforms expand considerably, concerns continue to surface about accuracy, legal privilege, and the environmental impact of AI-generated outputs. Though these technologies have benefitted some affordable housing practices, not all industry leaders have embraced these technologies at the same pace or for the same purposes.

Large Language Models
In the simplest terms, Large Language Models (LLMs) are complex programs that use a series of pattern-seeking algorithms to sift through vast collections of language-based data. As these algorithms parse the data, the program is “trained” to recognize patterns, a process known as “machine learning.” LLM users will send a request for information (a “query,” in technical parlance) to the program in plain human language, rather than employing complex search terms or manually collating information. And over time, as the program recognizes and synthesizes patterns, the LLM can provide answers to queries in simple human language, pulling from the datasets on which it has been trained.

Today, many professionals interact with LLM-based chat platforms, such as OpenAI’s ChatGPT, Google’s Gemini, and Microsoft’s Copilot. The proliferation of these platforms has led to near ubiquity in end-user contexts, with LLM integration being pushed to many enterprise productivity suites in recent years. In addition to direct usage of these products, users will commonly encounter LLM-enhanced chatbots when interacting with customer service platforms. 

In an affordable housing context, these LLM-enhanced programs have begun to proliferate, particularly within tenant service portals.

Jason Newman, vice president of asset management at Pennrose, is optimistic that Pennrose’s new “virtual agent” service, powered by an LLM-based chatbot, will lead to a more frictionless and efficient experience for everyone dealing with the mundane daily realities of the rental market – tenants, managers, and owners alike. “We want to be at a point where you can have a conversation with our information,” Newman says.

“We have it turned on at about 35 sites right now, and it’s an incredibly powerful platform. We’re still trying to work through exactly how to leverage it to its fullest extent,” he continues. “Say a tenant has an issue with a fixture, like the toilet in their unit,” Newman says. “The AI can automatically generate pathways [for the tenant] to self-correct, or it can directly escalate a catastrophic leak to our maintenance team. …It’s been really good on that front, and on managing call volume.”

Accuracy and Scalability
LLMs have become more sophisticated in recent years, with improvements in computing power and underlying algorithms. But the foundational technology of parsing large datasets is not new. For decades, the earliest predecessors of LLMs were used in academic contexts, notably in the field of linguistic corpus analysis. During their research, linguists often sift through an extensive repository of text (known as a corpus) to identify pattern-based changes in language over time. As the technology developed, researchers wishing to automate these methods would build their LLM algorithms from first principles, owing both to the lack of widely available LLMs and the need for accurate and context-specific output.

One of the principal breakthroughs of the last few years that has eased the widespread adoption of LLMs has been advancements allowing prompts and outputs to be formulated in plain language, rather than in code language or specialized search terms. However, not all LLMs are created equally, and for highly specialized fields, such as the affordable housing sector, potentially inaccurate outputs can pose a major concern to those who wish to seize upon the potential efficiencies of a more automated workflow.

These concerns about accuracy, along with the cost and scalability of third-party LLMs, have led some firms to build their own AI agents that can accomplish specific industry tasks. One such company is RibbonOS, a recently-launched market data analytics platform for affordable housing providers. “It’s our own AI agent that we built in-house. We tried ChatGPT, but this is not a scalable execution due to cost over time,” says Curvin Leatham, the company’s CEO. “Fine-tuning the output is the hard part,” he adds.

Leatham says that the outputs of Ribbon’s custom-built models—which also integrate Predictive Analytics—are already yielding exponentially more efficient (and accurate) results in their clients’ site analyses. “Over the past six months, we’ve been testing out the software in-house, and now it’s live. Some of our users have access to the integrated AI, which will look at the demographic tables or history tables and interpret and analyze what the data is telling you. It’ll also provide recommendations on next steps – what the data could mean for that particular area,” he says. “You can probably do about 40 percent of a market study through our AI software.”

Predictive Analytics
Aside from LLMs, Predictive Analytics (PAs) are the other primary AI technology that has seen buy-in from the affordable housing sector. PAs are like LLMs in their deployment of machine learning, but unlike LLMs, PA-based platforms synthesize numerical data, and their outputs are not commonly presented in plain human language. Standard outputs include cost projections and statistical analyses.

Like LLM technology, Predictive Analytics (PAs) are not novel. Anyone who has ever used Microsoft Excel’s “Flash Fill” feature to auto-populate a spreadsheet column has used a primitive form of PA. And like LLMs, PAs have become considerably more sophisticated in recent years with advancements in datacenter computing power, resulting in more robust platforms that can parse and visualize data in a single step.

This has led to the gradual adaptation of the technology in the housing sector, particularly as applied to site analyses, which factor datasets from a variety of geographic, demographic, and financial sources. That was the impetus for Leatham’s team to pursue a blend of LLM and PA technology in the development of RibbonOS, where PA-generated analytics are used as the basis of the LLM-generated language output, which, in turn, can be easily adapted by the user to suit their stylistic preferences.

“Our output is pretty beefy,” Leatham says. “Some of the reports that export from our systems are 30 to 40 pages, because it’s analyzing every table and demographic report. …We want to give you as much as possible, and you can just ignore whatever isn’t helpful.”

At Pennrose, the development practice has begun experimenting with PA-based reporting as they move through the project lifecycle. According to Chief Information Officer John DeSantis, the potential capabilities of PA have been promising thus far, but still require immense planning, supervision, and fine-tuning.

“On the asset management side, we’re working with consultants on how we could do more predictive analytics using AI; nothing is in production right now, but again, we’re still understanding what’s possible and which path we should take with AI,” DeSantis says.

DeSantis stresses that his team is doing robust work to ensure that any AI adoption won’t pose security risks. “AI doesn’t forget,” he says. “Once it knows that information, it’s never going to forget it.”

A Tool, Not A Replacement
The recent wave of popularity for AI has prompted concerns for many experts, with uncertainty around job displacement, cognitive impacts, and erosion of workplace skills. DeSantis stresses that Pennrose is not seeking AI functionality to replace its employees.

“I think the misconception is that it can replace everything. It’s going to make us more efficient in certain areas and make us more productive. But it’s not going to do everything all the time. …I still need human oversight,” he says.

Still, DeSantis says his team is excited about the positive possibilities of AI for team productivity. “We are starting to invest a lot of time and resources in where we can deploy AI agents to reduce or remove the manual, repetitive work that’s being done, especially on our accounting side every month. That’s a huge thing right now for us. …We’re not looking to replace anyone, but how can we better utilize their time in other areas, and have these [AI] agents work for us?”

Leatham recalls his days as an analyst, when jobs took days (or even weeks), and compares the improvement of AI functionality to another familiar moment in recent history.

“Think of banking back in the early days…You had to go to the bank to get cash, but then they invented something called the ATM,” he muses. “That’s what AI is allowing people to do. Instead of having to go and wait in line and fill out a slip, it’s like a data ATM, for people to do things more efficiently.”