Knowledge Work, AI, and the Real Blockers in Commercial Real Estate

What We Really Mean by “Knowledge Work”

There’s a lot of conversation lately about automating “knowledge work,” and it’s worth pausing to unpack what that even means. Most modern work involves someone with an understanding of a domain—legal, financial, operational—taking in information and applying judgment. A lawyer interprets documents through the lens of case law. An insurance underwriter evaluates risk based on policies, history, and regulatory constraints. A property accountant or AP specialist codes invoices by drawing on knowledge of GL structures, buildings, vendors, and how things really work inside their particular organization.

From a distance, these jobs look straightforward to automate. Machines can classify text, extract fields, cluster patterns. They can even appear to “understand.” But that word—understand—is doing a lot of heavy lifting. To understand something is to build an internal model of the world, a subjective mapping that lets you explain a concept to someone else who also understands it. Learning is the process of building that model. Intelligence is how fast something can build it. And judgment is simply what you do with that model when new information comes in.

AI systems do a version of this, but their understanding is entirely dependent on what they’ve been exposed to. That’s where the blockers show up, and those blockers are very different depending on the domain.

When Automation Is Blocked by Humans

In some fields, the blockers are societal. Even if the technology were fully capable of making a good judgment, the public simply isn’t comfortable with the idea. It doesn’t matter how accurate an algorithm might be—people don’t want a machine determining criminal sentencing, or diagnosing them without a doctor involved, or deciding who gets a job interview. In those domains, the limiting factor isn’t capability; it’s appetite. People want a human being in the loop.

In other domains, that discomfort just isn’t there. Insurance underwriting is a good example. The applicant doesn’t really care whether a human or a machine decides the risk score, as long as the pricing is fair and the policy is issued quickly. Commercial real estate falls squarely in that category as well. Vendors don’t care whether a human or a model codes their invoice. Property managers don’t care whether a human or a model flags fraud, or matches payee information, or triages a payment exception. They care about accuracy, timeliness, and predictability. The emotional stakes simply aren’t the same.

So if the public appetite for automation is there, and the tasks themselves are mechanical enough, why isn’t commercial real estate already fully automated?

The Real Issue Isn’t the Model

The short answer is: the models are capable, but the data isn’t accessible.

One of the reasons AI systems can pass the bar exam or the medical boards is that the training material for those domains—cases, textbooks, papers, reporting guidelines—is widely available on the public internet or in digital form. These models have absorbed decades of open, structured, richly labeled examples.

Commercial real estate is the opposite. Nothing about CRE operations is represented in the public pretraining data of these models. There is no publicly available corpus of property management agreements, invoice layouts, vendor rules, historical coding decisions, exception workflows, or GL structures. All of it lives inside private systems and idiosyncratic processes. It sits in Yardi, MRI, and RealPage databases; in Excel files stored on shared drives; in PDFs buried in email threads; in bookkeeping conventions that vary from property to property, portfolio to portfolio.

Why CRE Data Is So Hard to Learn From

These systems aren’t just private—they’re often heavily customized and tightly controlled. Gaining access to the right information isn’t a matter of pointing an embedding model at a folder. It’s a matter of negotiating access, building integrations, normalizing messy historical data, and accumulating labeled examples across a wide and constantly changing variety of formats. The problem isn’t that AI can’t extract a vendor name or detect fraud risk. The problem is that AI can’t learn an understanding of commercial real estate without the raw material—and the raw material isn’t lying around waiting to be scraped.

That’s why, when organizations in CRE talk about “building AI,” what they’re really signing up for is not model development. It’s years of infrastructure work: connecting to systems of record, mapping fields, resolving inconsistencies, maintaining pipelines, dealing with drift, and cleaning up messy inputs from vendors, properties, and accounting teams. The real investment isn’t the model. It’s the data gravity surrounding the model.

Where Build vs. Buy Gets Real

This is where the build-versus-buy question becomes more interesting. Prototyping an AI feature is easier than ever. Any team can spin up a proof-of-concept in a weekend if the data is already in the right shape. But in CRE, the data is not in the right shape. It’s not even in the right systems. Which means the real question shifts from “Can we build the AI?” to “Do we want to build the integrations, pipelines, relationships, and labeled datasets that give the AI something to understand?”

Most commercial real estate organizations don’t want to become data-infrastructure companies. Their core is asset performance, NOI, tenant experience, investor reporting—not training domain-specific models or maintaining ingestion pipelines. But unless an AI solution already has deep integrations, domain expertise, and years’ worth of labeled CRE-specific data, the automation will never get beyond the prototype phase.

The Only Blocker That Actually Matters

That’s the real blocker in this industry. Not public acceptance. Not model sophistication. Just the simple fact that understanding a domain requires exposure to it—and in commercial real estate, the exposure isn’t freely available.

Automation is absolutely possible in CRE, and the appetite for it is high. But the heavy lifting happens long before the model ever writes its first line of output. It happens in the unglamorous work of unlocking data from old systems, standardizing it, enriching it, and building the connective tissue that lets a model learn. That’s where the true investment lies. And that investment is what separates a proof-of-concept from a reliable system.

In the end, the question isn’t whether AI can automate knowledge work in real estate. It can. The question is whether a given organization wants to take on the data-infrastructure challenge required to make that automation meaningful—or whether it makes more sense to buy a solution that already has the understanding built in.