AI Comes for CRE Lending

A lot of ink has been spilled on AI's impact on the real estate industry. Usually, these analyses are focused on AI's role in either construction and development or building operations. We've written on each – specifically, thought experiments on what it might look like to build an AI-only real estate developer or a zero-employee property manager.

One of the most important actors in the real estate ecosystem, however, has largely been ignored: the lender.

But that is beginning to change. On one side, tech companies are increasingly building AI-enabled tools for commercial real estate lenders. And a new crop of tech-enabled firms are attempting to build CRE lenders from the ground up using AI from the start. And unlike the implications of AI-enabled operations and construction, what it might mean for commercial lending to be AI-powered has largely been overlooked.

Today's letter will shine some light on this less-known corner of AI in real estate. Specifically, we'll tackle:

• The opportunities and challenges of AI in lending;
• AI applications and players in CRE lending;
• What AI-enabled lending means for the real estate world.

Let's dig in.

Caution and Potential

Of all the players in the real estate ecosystem, lenders are perhaps the most obvious case for the broad application of generative AI. Many of the nuanced and challenging things that block widespread use of AI for real estate developers and property managers — think construction document production, zoning analysis, and handling challenging tenants – don't exist for lenders.

If anything, CRE lenders generally have more defined buy boxes and repeatable deal processes that work nicely with AI. And borrowers are usually providing their lenders with structured data on deals, enabling more straightforward automated deal evaluation against known, quantifiable criteria. Similar structured data can be created on the servicing side, particularly if baked into covenants from day one. In theory, AI should be able to do much of the lender's job and dramatically change the lending market.

But things in real estate are rarely so easy.

Historically, lenders have been the most conservative players in the real estate ecosystem. And that makes sense: whereas equity investors capture the upside of a real estate deal, at best a lender will see the return of their capital plus interest. This dynamic has imbued lenders with a culture of moving slowly and carefully.

"There's a lingering 2023-era notion that ChatGPT hallucinates and makes everything up," said Yaakov Zar, founder and CEO of Lev, a company providing SaaS tools to debt brokers and real estate sponsors to manage lender relationships. "These are very large transactions with a pretty significant risk-reward balance. If I'm making a $10 million loan, I need to CYA with any research I'm doing."

Zar adds that many bank lenders are prevented from using AI tools by regulation. "They can't even use GPT to summarize a document, which is insane," he adds. "There's a huge regulatory blocker here."

CRE lending is also very much a relationship business – much more so than development or operations. And to the extent lending happens beyond the domain of tight personal relationships, debt brokers typically sit in the middle of those transactions, creating a principal-agent barrier to lenders seeing the benefits of AI adoption.

But change is on the horizon as AI seeps into every corner of real estate.

Introducing Asimov Capital

Last year we introduced Asimov Partners and Asimov Management, real estate developer and property manager (respectively) with zero employees between them. Today we'll introduce Asimov Capital, a CRE lender with zero employees.

Asimov's deal sourcing process is quite easy: it's an online form. Unlike a developer or property manager, this Asimov doesn't need to go out of its way to find opportunities – if Asimov's capital is competitive, owners and brokers will find it. And with technology driving operating costs to ~zero, we believe our capital will be very competitive.

While Asimov's deal sourcing is fairly low-tech, evaluating those deals is where AI comes into play. Typically, a lender's evaluation process happens in two steps: first, a quick "back of the envelope" to see whether a deal might be viable. If the deal passes muster, the lender will then embark on a more thorough underwriting process to produce a term sheet.

Since Asimov doesn't need to worry about wasting time, we'll collapse these two phases into one automated underwriting step.

This isn't a totally theoretical exercise; some real-world lenders have begun using AI to simplify their underwriting processes. "We use AI to do a lot of things that were cumbersome or time-consuming before," explains Guillermo Sanchez, co-founder and CTO of private multifamily lender Ease Capital. “AI is great at the standardization of things. We take an OM and extract everything we find useful for a loan request," continues Sanchez. "We use it to do standardization of rent rolls and operating statements, saving analyst time."

Fortunately for lenders without an AI-savvy CTO, a number of third-party tools have emerged in recent years to make this possible without writing custom code. Blooma and Archer are two examples of technology companies building AI-powered software to automate the underwriting process. Archer, for instance, promises "address to complete underwriting in 15 minutes." Lenders are prompted to upload a rent roll and T12, and Archer will automatically parse those files and pull in relevant comps.

While Archer and Blooma offer similar feature sets, Asimov Capital will use Blooma given its API functionality – a critical feature when we have no humans to press the buttons.

Of course, we're not going to finalize underwriting without some double-checking – we want Asimov's robot brain to embody some of a human CRE lender's natural conservativism. And while LLM hallucinations are much less common today than they were 24 months ago, they do still happen. “You don’t want to have hallucinations on line items of an operating statement," notes Sanchez.

So we'll take two steps to ensure Asimov's underwriting is within the bounds of reality:

• A retrieval-automated generation (RAG) system pulling information like comps from a defined list of sources, and
• A separate "checker" AI that validates underwriting inputs versus the provided information.

These are similar to the steps Lev takes to validate AI outputs. "The power of agentic systems is that they can check each other and themselves," explains Zar. "When we build an agentic system, we make sure it’s only referencing the materials we provide, and we have a checker looking at it versus the original input."

Done right, AI-powered underwriting can do more than just speed up the process. “We believe we can pull up a lot of traditional underwriting complexity that would happen further down the process," explains Sanchez. "We can pull up front and have a more informed view by the time we give you your quote.  So we’re not re-trading you down the process.”

“There are some soft KYC checks we can do using AI," Sanchez continued. Has [the applicant] had a lawsuit in the last 6 months? Crime on property? That happens automatically as a first pass check. There are a million little things we can automate and present in a single pane of glass that let us direct our efforts in the deal pipeline.”

Moving these capabilities further up the funnel means that AI-enabled lenders like Asimov (and Ease) could hypothetically produce a term sheet with a higher certainty of close than a human-powered lender which has to push those additional checks further down the funnel given the time and effort involved.

Diligencing and Servicing

After a term sheet is negotiated and accepted, lenders typically move onto the third phase of their process: conducting due diligence on the deal.

For most lenders, this phase is still very handmade. “At the last stage [diligence] you’re so far down the pipe, it’s worth it to hand off to a closing team to deal with," noted Zar. That is, the number of deals making it to the diligence phase is so small relative to the risk of getting something wrong that lenders are loath to hand much off to computers.

But Asimov Capital has no human employees, so we have no choice but to hand this step over to the machines. And fortunately, there are third-party tools at the ready. AI-powered property diligence software like Prophia, for instance, covers everything from leases to contracts to rent rolls to encumbrances. While it appears to be primarily designed for acquisitions teams, there's substantial overlap with what a lender would want to see. Lightbox has more lender-focused tools for streamlining appraisals and managing collateral, although it's not totally clear how automated those tools are. Unlike Prophia, Lightbox does appear to have an API, so Asimov will use that to automate as many diligence steps as possible.

Where humans are needed in the diligence process – well, Thumbtack now has an API, enabling Asimov's robot brain to initiate all sorts of chaos by sending appraisers and inspectors to sites at will.

Once our loan is diligenced and closed, Asimov will need to service it – an area where Guillermo Sanchez and Ease Capital see a lot of potential.

“25 loans per person on servicing side is traditional," said Sanchez. "But can we use technology to multiply that by a significant number?”

"For each deal we have an 'AI deal room,' and all documents get put in here. We can ask questions of that deal, we can look at performance vs expectations. As we get updated rent rolls, we can see what the gap in performance is."

Sanchez emphasizes that he's not looking to rebuild or replace an asset management tool like RealInsight.  "We want to sit in the middle between servicers and RealInsight so we can monitor that against our own loans and performance," he explains. "We’re not trying to rebuild those tools, but we’re trying to integrate in a way that allows us to be efficient.”

“It’s crazy to me that people don’t know a loan is going to go south until the last minute. There’s often a trail that would let them have some insight into that, and the ability to continuously ask questions about a deal is super valuable."

AI also enables more intelligent management of loan agreements that may differ in their terms. Natural language processing, for instance, can extract key terms from loan agreements (DSCR thresholds, reporting deadlines) and automatically monitor compliance.

“In this environment, lenders often have to get creative to close certain loans," said Sanchez. "Managing the complexity that creativity creates [in varying covenants and terms] will be a key problem, and that’s where people get most excited when they talk to us on the asset management side.”

Of course, improving loan servicing isn't just about better organizing and interrogating data – entirely new types of data are available as well. Computer vision and data integration, for instance, can update collateral values using sales comps, market rents, and even satellite imagery or drone footage of construction sites. AI tools can then generate standardized reports (e.g., CCAR, stress tests) from raw servicing data with fewer errors and less manual effort.

But what doesn't yet exist in the wild is a replacement for a lender's brain. All the tools that exist today are designed to cue up the right data at the right time in front of a human servicer. The actual decision of what to do in the event of a breach of covenant or default is left to a human being. In practice, lenders often have many tools at their disposal – acceleration of repayment or foreclosure, for instance – that they don't actually end up using when a breach happens. In most cases, those tools are simply there to force the borrower to have a conversation with the lender and explain what's going on.

In other words, a maximalist approach to enforcing covenants and repossessing properties is loan shark behavior. So any AI replacement of a human lender would need to be trained on situational nuance. How deep is the lender's relationship with the borrower? Do they have other deals together? Does the borrower appear to be acting in good faith? Is there hope that the borrower can turn things around? How much of a hassle would it be to take back the keys? Is the lender facing regulatory pressure?

As far as I'm aware, nobody is solving these problems – nor should they be, as the softer aspects of relationship management and decision-making around special servicing are held tightly by lenders and unlikely to be handed over to machines any time soon. But the rules of our thought experiment are the rules, so we'll need to spend some time training Claude on when to let things slide... and when to take the keys.

Optionality

As a real estate GP borrowing money in good faith for projects, it's hard not to see the growing role of AI and automation as a positive thing. In a competitive, commoditized market, lower overhead and administrative costs should largely be passed through to customers – in this case, borrowers.

More importantly, perhaps, better automation – particularly when it comes to data – will break down the barriers that have long constrained many lenders to operate within narrow geographic mandates. Suddenly, a New York-based lender can understand and monitor a property in Michigan or California in a way they never could before. In theory, this means any given borrower has more lenders with the ability to submit an LOI on his or her deal.

But this reality is still a long way off. Most lenders, and particularly the local and community banks with the cheapest capital, are years away from implementing AI in a meaningful way across their firms. Many local banks still struggle with the basics of online and mobile banking, 2010-era technologies. The chances that they suddenly emulate our Asimov Capital thought experiment seem slim. Of course, this opens an opportunity for private lenders like Ease to fill the gap.

"Most [lenders] don't have in house talent to build [or] integrate and the outside vendors often build complicated solutions that don't solve for how any one particular organization works," said Sanchez.

"Real estate is inherently a meatspace business. Unless you have already embraced technology, the jump to AI is pretty big if there isn't a clear integration."

-Brad Hargreaves